Meta-essay

On October 2nd 2026 I was invited to give a talk for TEDx Aston University.  It was fantastic to be asked and just the right amount of terrifying.  Even as someone who is fairly used to public speaking, I got a little freaked.   The TED format is quite unusual as both a genre and in its length, and I elected to speak without notes or slides, which is typical but not absolutely necessary for a TED talk.  I was supported in the talk by Elinor Vettriano who provided speaker coaching for the event and I spoke alongside Helen Higson, Dileep Marway, Avni Ahuja, Kamese Davis and Sharon Thompson all of whom gave well-delivered, interesting and sometimes personal talks – check their talks out when they appear up on the TED website.

I think when I was invited the organisers may have wanted me to deliver a talk on forensic linguistic casework, but I find this content difficult and often depressing, and it seemed a poor fit for an upbeat TED event. It is also well covered in the Writing Wrongs podcast I co-host with Nicci Macleod. Go there if you want that. Instead, I chose to spoke about the development of the essay as an attempt to improve and clarifying thinking, and how this contrasts with current University use of the essay as an assessment, and I also wanted to discuss the contemporary crisis for written assessments in the age of AI.  

One aspect of my talk focused on the etymology of the word ‘essay’ from the French ‘essayer’, a try, or an attempt.  In preparing my talk I didn’t write it out and memorise the text.  I wrote and re-wrote my talk as a series of bullet points that I wanted to talk around.  I was happy with my attempt on the night.  Naturally I missed some of the points I meant to cover, and I didn’t land all of my lines, but overall the talk went well.  I am pleased with the outcome and very pleased to have been asked to do it.  

The spirit of my talk though was that the thinking should never stop and – so I came home and decided to have another go. This time, obviously 🤷‍♂️, I wrote my talk as an essay. As you will read below, it’s not a very academic essay. I’ve not researched the facts in the way I would if writing an academic paper – for my original talk I looked up some dates etc. but I’ve not thoroughly re-read Montaigne or Locke – so experts in those writers may have to give me some slack. I hope I’ve not misrepresented their thinking, but I’m aware I’m being broad brush and using them for rhetorical effect. I’ve not formally referenced my essay either, but I hope there’s enough information within what I’ve written for an interested reader to find the people and texts I mention.

So, this attempt at the The Life and Death of the Essay is what it is – another attempt.  It may not be the final attempt.  It is longer and some of the arguments are filled out, and hopefully I’ve landed more rhetorical lines than in the TEDx version.  I’ve still written it as if for a talk – it’s come out in a kind of hybrid register between a speech and a written essay.  I guess it’s a bit more formal than theTEDx version, like a written speech. 

You’ll find a bit more here than in the live event or when my talk appears on YouTube.

If this is too long for you – these days it would count as ‘a long read’ — then the TEDx talk is a shorter attempt of what I’ve been thinking about.

Do leave comments and criticism – you’ve missed the opportunity to throw fruit at me this year, but if that’s what you’d like to do, perhaps TEDx will invite me back for another talk another time.

 

The Life and Death of the Essay

  • How good are you at thinking?
  • How do you know how good you are at thinking?
  • How would you improve your ability to think?

These were questions that were troubling Michel de Montaigne in 1570.

To answer these questions Montaigne invented a new modern technology for improving thinking – he invented ‘the essay’. 

In the ten years after 1570, Montaigne shut himself away in a tower, in his chateau, in the Dordogne, in rural France, and wrote more than 100 essays.  

He wrote on a world of topics.  He wrote and rewrote about religion and the beginnings of science, he wrote and rewrote about ethics and politics, he wrote and rewrote about education.  Montaigne’s insight was that by writing an essay you are placing your thinking outside of your head. You are making your own thought the object of study so that you can improve the rigour and creativity of your thinking.  In spite of the breadth of content in his essays, when he published his Essais in 1580 he thus wrote that ‘I am myself the matter of this book’.  

Of course, if you have written an essay you can also pass it around, invite others into your thoughts.  When you read an essay you are able to get inside someone else’s head, you are able to reconstruct and try on their thoughts – it’s an invitation to critical thinking.  Both writing and reading essays can be a training in critical thinking.  

The word ‘Essais’ comes from the French verb ‘essayer’ – ‘to try’. Montaigne’s Essais  were ‘try-outs’, ‘attempts’ to describe aspects of the human condition and solve human problems.  It is one of the most important aspects of his essays that these attempts at thinking were never finished.  Montaigne’s Essais were first published in 1580, then again in 1588, and then again 1592.  When you look at the later editions of the Essais they are not re-writes in the contemporary sense.  The earlier versions are often left intact, but now Montaigne’s new paragraphs reflect on what he said before, sometimes overturning his earlier thinking, sometimes reinforcing it, always attempting to improve that thinking.  For Montaigne the thinking undoubtably continued beyond  his final edition in 1592 – and only stopped when he died in 1594.  For Montaigne the thinking only stopped when he was dead.

Montaigne’s Essais marked the start of a golden age for the essay that lasts 200 years before it starts to go wrong.  The essay was a massively important invention that powered the scientific revolution and the enlightenment, the age of reason, and the birth of modern science. 

For example, in 1988 when I was an undergraduate philosophy student, the very first essay I was asked to write was on John Locke’s Essay Concerning Human Understanding.  Locke, I was told was one of the greatest of all English thinkers. He didn’t get everything right. For example, he had despicable views on slavery, but on the other hand his political philosophy was the basis of the American constitution, which as we’ve seen, protected democracy in the United States for nearly 250 years.  

In the Essay Concerning Human Understanding Locke’s main interest is how we know something, how we know anything at all.  And his answer can summed up in his adage that “No man’s knowledge can go beyond his experience.”Locke is an empiricist.  He’s quite an extreme empiricist.  In an era before science he laid the foundations saying that the only things we can truly know are from our senses, our sight, hearing, taste, and touch. So for example, in terms of the human mind he was what we would now call a behaviourist.  We can only know another’s behaviours, not that they have inner states of mind – thoughts and emotions of their own.

Locke first started writing and circulating drafts of his Essay Concerning Human Understanding from 1671. And it took him 18 years of attempts and tries before he was ready to publish in 1689. And then he had another version, another attempt published in 1694, and another in 1695, another in 1700  and a final attempt published in 1706, two years after his death.  Attempt after attempt, 18 years of attempts before he published, and then a further five publications of the essay published in the next 17 years. And for Locke the trying, the thinking only finished when he was dead.

In the middle of the Essay Concerning Human Understanding there’s a great passage where John Locke, one of England’s greatest ever thinkers, gets taken in by a talking parrot.

The parrot belongs to the Prince of Nassau and is a truly impressive bird. It is said to be able to hold conversations.  If you ask it where it’s from it will tell you that it is from Brazil, it will also tell you it’s owned by a prince, and so on.  Locke uses this story to draw a distinction  between ‘a man’ and ‘a person’.  As a man, I am my animal, biological self.  As a person I am a rational being, I have a mind.  If a parrot can produce convincing strings of words, then it is a person – it has a mind.  This is a reflection of Locke’s  empiricist, behaviourist perspective that devalues inner states of others and focusses on empirical observation. The speech sounds are the evidence of a mind.

Obviously this is nuts. 

The parrot  doesn’t know what ‘Brazil’ means, it doesn’t know that the sounds are speech sounds, that these form a word with an external referent.  The parrot does not understand the social and conceptual superstructure of ideas that leads to labelling regions of the planet, to maps and lines on a map that mark out territories as countries with borders.  The parrot simply responds to the stimulus, to the prompt, Where are you from? with the response of convincing, grammatically correct but empty speech sounds, that it, the parrot doesn’t even know amount to speech.  To the human listener this is words and grammar, but for the parrot it is without meaning. Call and response. The parrot is essentially trained to extrude meaningless but relevant well-formed text.

But not for John Locke, one of England’s greatest ever thinkers.  

Let’s move forward from his Essay on Human Understanding to the beginning of the end for the essay, one hundred years later.  

It is now 1791 and the scholars at the University of Cambridge are about to perform a dastardly act. They are about to poison the essay as a mechanism to work out and improve thought, as conceived of by Montaigne and used by great thinkers like Locke.

In 1791 the University of Cambridge invents the assessed essay.  They introduce the unseen essay exam.  

They have their reasons.  Previously students were examined by oral ‘disputations’ judged arguments and debates between students.  These are time consuming assessments to run.  Also, Cambridge was the centre for mathematical learning and it was becoming apparent that in maths, oral argument was perhaps harder to judge, and that a written proof might be better.  The introduction of the written essay exam proved massively efficient and enabled the university to grow. There was now mass higher education – ‘matriculations’, graduations, tripled in the next 40 years from about 150 students a year to 450 a year. This was revolutionary for higher education in the same way as since I first enrolled in University, enrollments have also tripled from about 760,000 to 2.8million students a year in the UK. But Cambridge’s introduction of the assessed essay is sheer vandalism to the idea of the essay.

The idea of the assessed essay is inimical to the very idea of the essay in two specific ways.

First, it changes the primary purpose of the essay.  The primary purpose since Montaigne is the attempt to improve thinking about interesting problems. Now the primary purpose is to achieve the highest ranking, the highest grade. At first it seems that maybe these two things can be achieved together, but they can and do become separate and the highest grade becomes the priority.

Second, the assessed essay introduces the idea of a due date or an exam date, and this poisons the idea of an essay as an attempt.  It changes the essay from a process to a product. From an a ‘essayer’ to an ‘objet’.  What student, after the point of submission, continues their thinking on how to improve that specific essay?  We all of us who have been students know how rapidly we move on to the next assessment, and the next, and immediately forget the just submitted work. 

In spite of the University of Cambridge’s best efforts, at this stage the essay is not killed off, it is not dead, but it is injured and badly deformed by their actions. The death of the essay takes another two hundred years. We move from the 1790s to the late 1990s.

The 1990s see the invention of the internet, the world wide web, and by the year 2000 this new technology is available across most western universities. In turn the internet enables the invention of two wholly new forms of essay writing that world has not seen before. The internet enables the invention of two new forms of plagiarised essay.

First, there is the simple, free form of internet plagiarism where you find a chunk of text online that is useful in your essay, and you copy and paste it and then submit it as your own words. If you get really lucky, some other student at another university has written an entire essay on the topic you’ve been given, and posted it online, and you are able to cut and paste it intact – otherwise there’s always Wikipedia, which was conveniently launched in 2001.

Second, there is the invention of the essay mill, where you might pay a graduate student from across the world to write your essay for you. Commissioned essay writing started with bright, poorer students from the developing world, and quickly got taken over by organised crime.

Outsourcing the writing of your essay, or copying and pasting informative, well-formed text to answer your question, are both creative intelligent solutions to efficiently achieve the primary purpose of the assessed essay – that is to achieve the highest possible grade. But of course it completely detaches the writing from the thinking. From the year 2000 onwards thinking is no longer required to create a high scoring essay.

For Montaigne the thinking only stopped when he was dead. For Locke the thinking only stopped when he was dead. Nevermind the parrot, it is now the essay that is dead. It is deceased. It has cease to exist. It is no more. It is completely dead.

This throws universities into a panic. The primary and most efficient form of assessment for universities is now dead, but because universities are addicted to written assessments, so the essay lives on in zombie form.  The zombie essay can be indistinguishable from the truly human essay and can hide amongst the others. Unthinking, sometimes lurching, giving away uncanny clues to its lack of thinking, and generally infecting the student population.  Students who want to learn and understand that thinking is hard and must be practised, become disillusioned, aware that these zombie essays often achieve higher grades than their own. And it’s true students seduced by the zombie essays can be awarded high degree classifications through their dabbling in the occult. After all that’s the point of assessment – to get the best possible degree, the highest possible grade.

So Universities face a decision point, they face a choice.  Either they can examine their practices, reduce their addiction to written assessments that can be completed without thinking, or, they can invest in the science of zombie detection. They can go down the forensic route, the forensic linguistic route, to distinguish between the zombie essay and the essay produced by hard human thinking.

UK Universities (through an organisation called JISC) initially went down both routes.  For six years between 2002 and 2008 they invested in a new Plagiarism Advisory Service.  This had two initial remits.  The advisory portion provided thinking and advice to universities on how to reform assessments, how to design plagiarism resistant assessments in the age of the internet, and this produced lots of good creative ideas and advice that has developed into interest in authentic assessment today. But today this is a relatively poorly funded niche interest. The Plagiarism Advisory Service also evaluated forensic plagiarism detection services and allowed for the new companies offering these services to bid for a contract. This contract gave all UK universities three years of free access, after which each university would have to pay for such services itself. That contract was won by iParadigms, who offered the TurnItIn system.

I was a young academic in the University of Leicester in 2003, and although I was junior, I was also a  forensic linguist, and so I was co-opted onto a university committee that was asked to make the decision as to whether Leicester should adopt this new plagiarism detection software. To the surprise of some of my colleagues, I was the only member of the committee who was arguing agains adoption of TurnItInt. I remember saying that we didn’t want to make assessment more adversarial, more like the courts, that we would start seeing lawyers in disciplinary hearings, and that above all introducing these forensic techniques into universities would damage the relationship between teacher and learner. But I was junior and other views prevailed.

As an aside we all should beware free or unsustainably cheap software that you might become dependent on. After the first free years, universities had to start paying themselves. There’s no very accurate figures, but the estimate is that UK universities now pay between £10M and £20M every year subscribing to plagiarism detection software. This stands as a warning to those accepting free or heavily discounted AI systems into your business or university. In the rush not to be left behind we are missing that these are unfeasible cheap and so the technology companies don’t go bust, prices will have to soar.

And thinking of AI brings us to the present.  It was in just 2022 that ChatGPT was released to the public and the world became aware of Large Language Models (or LLMs) – and a new age of panic for universities started.

Let me be clear.  My view is the assessed essay died in 2002. The current panic about AI, about the zombie essay in the age of AI, is that this new technology is just revealing that the assessed essay has long since been dead. It wasn’t AI that killed the essay, it was the internet that enabled the efficient production of thinking-free essays. LLMs simply reveal that that the mechanical ability to string words together does not amount to thinking. Producing essays without thought is now just a lot more efficient. We no longer have to go to the trouble of all that internet searching, or finding an essay mill that we can trust. We no longer have to work out how to circumvent the plagiarism detectors. (Generally, that is not that hard, any forensic linguist can provide you with tips!).

Just briefly about LLMs: they are a new and very powerful technology and it’s true that any sufficiently powerful technology will be perceived as magic. Don’t be bewitched by the hype. LLMs are essentially only next word prediction machines. They are essentially very sophisticated auto-complete algorithms. They only differ from prior versions of auto-complete in what they have taken in as the previous words that they use in order to make their prediction of the next word. The objective for most LLMs is that they will have absorbed every previous string of words ever produced and written down by a human being. They take in and process all of the language, then the user produces a little bit more language that they take in and process – ‘the prompt’. And then the LLM maths predicts what the next word should be, and the next one, and the next one. That’s it. One of my favourite linguists to read, who writes about Large Language Models, is Emily Bender form the University of Washington. She was an author on the article that coined the term “stochastic parrot” to describe LLMs. Stocahstic just means probabilistic, and she also uses the term probabilistic Synthetic Text Extrusion Machines (STEMs for short).

LLMs just shows us that the mechanical ability to string words together does not amount to thinking.  It is well within their capability to get them to write you an essay about Brazil. But they know no more about Brazil than Locke’s parrot. The LLM doesn’t know what ‘Brazil’ means, it just predicts the words that would make an essay about Brazil, but with no external referent beyond language. The LLM does not understand the social and conceptual superstructure of ideas that leads to labelling regions of the planet, it’s just absorbed the text and predicts which words could be used to express those ideas.

LLMs are just a more efficient means of producing essays without thinking.

So again universities face a decision point, a choice.  We can go down the forensic detection route, or we can reform assessments to ensure that thinking is required to get a good grade.  

Detecting LLM produced text in student work is often portrayed as a tough problem.  And framed that way, as detecting LLM produced text, it is tough and probably a futile endeavour. You are aiming for a moving target – as LLMs improve, you are entering into an arms race with big tech.  But if you ask a forensic linguist, the LLM plagiarism problem can be an essay problem to solve. You should not frame the problem as a general categorisation task with human produced language in one set of texts, and AI produced language in another set of texts. This is because human produced language is such a large and varied set of texts, and LLM produced language draws on those texts and produces a text that is a contextually-tuned, averaged version of the human language.

If you ask a forensic linguist, this forensic linguist, you should start with a particular, specific human, and then model their language style. You could take a first-year student, for example, writing in a room without a computer, and you start with that text. Later if they submit a text produced with LLM assistance, or wholly written by LLM, it is easy to perform an authorship analysis that measures that shift in style. At the Aston Institute for Forensic Linguistics, we have already developed Aston-made software that can easily perform this kind of authorship analysis. These are the systems that we apply to criminal cases, cases involving murderers, terrorists or paedophiles writing on the dark web. Our evidence is often accepted in court and can help convict the guilty or acquit the innocent. Naturally to apply this approach to essay writing students you would have to produce a language model for every student that enrolled, and this would be a time consuming system to set up, and expensive. But if you pay me £10M – £20M a year I can do that for you. I can produce software and a governance system to achieve this individual surveillance of every enrolled students’ language style.

Who knows such a system might have wider uses – the police could serve a warrant to access the language model if they suspect a student (or ex-student) of some crime or some online harm. It would become a library of human language styles much like a DNA database or fingerprint database.

But even beyond the ethics and civil liberties aspects of this, even beyond the issue of introducing students to the university with a declaration of such an adversarial relationship, this is not what we should be doing.

The forensic LLM detection route fails to recognise that the problem is not how the text was technically produced – with a pen, or a word-processor, or an LLM – the problem is the separation between writing and thinking.  When our students graduate, they will for the next 40+ years of their careers be using all the technologies available to them to progress in their lives, and we hope they will use their creativity and the available technologies to make a positive mark on the world. We should not be setting them written assessments and essays and demanding that they write them like it’s the year 1990. This is no preparation or foundation for their contribution to the rest of the 21st century.

Universities should make the harder choice the choice to reform the entire process of assessment. We should find ways to enable students to use technologies and complete them in such a way that that thinking and learning is required to score highly. Universities should ensure that it is the creative and rigorous thinking that is assessed. This is a hard problem, but academics and universities round the world are already tackling it. There are already suggestions and ideas ready for investment. This is where our hard cash and hard thinking should be directed.

As a post-script, a few sentences ago I said that LLMs show us that the mechanical ability to string words together does not amount to thinking.  AI advocates would have it that these Synthetic Text Extrusion Machines (STEMs) are in fact thinking machines.  Tim Altman, founder of OpenAI, responded to Emily Bender about three years ago with a tweet that said “I am a stochastic parrot and so r u”.  This is nonsense – LLMs know no more about Brazil than Locke’s parrot. 

Locke’s parrot is instructive here in our response to what is claimed as artificial intelligence.  We should ask the question Why did John Locke get taken in by a talking parrot? and the answer is interesting.  Although John Locke may be one of the greatest thinkers England has ever produced, he is also human.  It is a function of our humanity as meaning-making creatures that we like to project meaning and meaning-making onto the world.  We create gods and monsters in fire and shadow, we speak to our plants and our cars, we endow non-human creatures and inanimate objects with personalities.  Where another creature or a machine or an algorithm produces language, this tendency is amplified. I say please and thank you to my digital assistant, for fun, and because it feels like the interaction demands it. We project a mind onto a talking parrot because the way we see the world requires that all language producers must have a mind, must be a person. This has been usefully referred to as an asymmetric interaction. The provision of meaning in th talk between the two interactants interaction is one-sided. And the same is true of our interactions with LLMs, with these Synthetic Text Extrusion Machines. STEMs need our humanity to appear to have minds.

And this is the thought I finish with – STEM needs the humanities.

The real crisis in today’s universities is not the death of the essay but the potential death of difficult, human, critical thinking. In universities the kind of thinking that we need in the age of AI is centred in arts, humanities and social science faculties. The disciplines that traditionally write essays. We see the death of thinking institutionally reflected in the devaluing, the defunding, and the death of these faculties in the across universities in the UK. But to think about this crisis properly would require another talk, another attempt, another essay.

User Centred Explanation Design in AI produced forensic evidence 

This post was first published on the Aston Institute for Forensic Linguistics blog 27 November 2025

In 1986 Donald Norman and Stephen Draper published the influential edited volume User Centred System Design, a collection of essays that set the agenda for software development and the field of Human-Computer Interaction throughout the 1990s and into the twenty-first century.  In his own chapter in the volume, Norman writes that “user-centred design emphasizes that the purpose of the system is to serve the user not to use a specific technology, not to be an elegant piece of programming.” (Norman & Draper, 1986, p61). As in many forensic sciences, in forensic authorship analysis there has been a rush to develop and adopt the latest specific machine learning technologies such as transformer deep learning architectures, adversarial learning, and Large Language Models (LLMs), which are often colloquially referred to as Artificial Intelligence (AI) systems.  These AI systems process and build models based in enormous amounts of data, often finding associations between and within enormous collections of variables, the scale of which means that these associations are sometimes uninterpretable by traditional statistical means and are certainly beyond the comprehension of human users of these systems.  The demonstrable power and accuracy of these systems in experimental datasets is often only matched by their opacity, leading the human analyst to bewilderment at specific outcomes, and crucially with no way of understanding how the outcome was arrived at.  In the excitement of new and powerful technological potential, Norman’s principle of serving the user seems to have been left behind.

In most jurisdictions AI produced forensic evidence is current not admitted in court cases.  One dilemma that forensic applications of AI systems can face, is illustrated by considering two sets of studies:

  • Basu et al (2022), Basu et al (2023) and Bali et al (2024) set out a series of mock-judge and mock-jury studies making use of an AI forensic speech science system known as E3FS3 (Weber et al, 2022), a forensic-voice-comparison system based on state-of-the-art automatic-speaker-recognition technology. E3FS3 was developed by to comply with modern principles for evaluation of evidence, i.e., that methods be transparent and reproducible, logically correct, resistant to cognitive bias, and calibrated and validated under casework conditions (Morrison, 2022).  The Basu et al (2022, 2023) and Bali et al (2024) studies compared the perceived strength of evidence of the human listeners vs the strength of evidence output by the system and did so with the human listeners acting alone (as a mock individual judges) and working in groups (as mock jurors).  The results of these studies are startling:  E3FS3outperformed all individual listeners and 78% of the groups of collaborating listeners.  When the output of the E3FS3 system was provided to assist human listeners in making their judgement, the performance of listeners provided with the output of the forensic-voice-comparison system was better than the human performance alone – the system improved human performance; however, the performance of listeners provided with the output of the forensic-voice-comparison system was worse than that of the voice comparison system without human judgements – the introduction of human listeners degraded the E3FS3 system performance.  These results suggest that an expert witness in court would be required to present, explain and defend such a system in its assessment of the evidence, but that the human jury (or other decision-makers) should not be allowed to listen to the evidence themselves and make their own judgements.  They should accept the evaluation of a well-validated system on the basis of the validation alone.   The system once validated (and calibrated to a particular case) cannot be questioned or confronted – in its use in court it is an ipse dixit machine.
  • A second set of studies relates to challenges put forward by the PAN authorship analysis competitions (https://pan.webis.de).   Since the introduction of authorship verification tasks (of various designs) to the PAN competitions, AI and complex computational approaches appeared to have solved the problem such that by 2021 all results in that year’s verification task competition exceeded 93% accuracy (Stamatatos et al., 2022). However, for the 2022 competition, a challenge was introduced relating to a more forensically realistic problem, that of cross-genre or cross-discourse-type analysis drawn from the 100 Idiolects Corpus (Kredens, Heini and Pezik, 2021).  The results of the PAN2022 competition were that the best-in-class systems and models, when applied to the 100 Idiolects Corpus, were reduced in accuracy to rates close to chance.  Stamatatos et al (2022, p11) write that the previous PAN competitions gave “the false impression that authorship verification is an almost solved problem”.  For linguists generally and forensic linguists specifically, the fact that cross-genre analysis would present real challenges to the computational approaches was unsurprising.  For the computational researchers, genre homogeneity / heterogeneity could be characterised as an unrecognised source of bias in the training sets.  In applied AI work it is well acknowledged that understanding such biases are crucial, but this had not come to the fore in the previous PAN authorship competitions.  It is easy to envisage a real forensic case where an advanced authorship analysis system was applied with the validation that it was at least 93% accurate – but where, if the case happened to involve a cross-genre analysis this would create a very real risk of a miscarriage of justice.  This situations demonstrates one risk of relying on system outputs which cannot be cross examined by human experts.  Human expertise and understanding are essential not only for the development of forensic evidence AI systems, but also for their use in specific individual cases in court.

In law, hearsay is excluded from US and European courts on the basis of the 6th Amendment defendant’s right to confront a witness testifying against them, and more or less equivalently, Article 6(3)(d) of the European Convention on Human Rights (ECHR) provides for the right of the accused ‘to examine or have examined witnesses against him’.   Jurs and DeVito (2024) argue strongly that AI systems that lack transparency and explanation and so breach these rights to  confrontation should not be admissible in court. Their proposal to avoid this risk is that “that only sufficiently transparent and explainable systems (open-box AI systems) should be permissible as evidence in courtrooms today” (p597).  They go on to say that a system is explainable “if, when answering queries, it presents information to the user that provides a qualitative understanding of the connection between the input and the output (p614)”.  

The answer to the dilemma of human involvement being both necessary and reducing the efficacy of AI systems is case-by-case explanation of the connection between input and output.  This doesn’t have to be a naïve rejection of the complexity of science.  Ypma et al, 2023 argue that “Triers-of-fact do not need to understand the details of forensic evidence evaluation – they need to be able to trust the experts and the systems they use.” (p5). They point to the complexity of a glass fragment analysis, suggesting a kind of chain of understanding – arguing convincingly that the jury don’t need to know the relationship between the chemical composition of glass fragments and its reflective properties, so long as someone does.

As Ypma et al, (2023) caution, this view of complexity in  forensic science as cannot account for AI generated forensic evidence where a ‘black box’ system is at the heart of the analysis, as this breaks the chain of explanation.

In this instance the development of the technique will involve training a system in an authorship analysis task.  This is likely to involve use of the latest specific technology of LLMs, where input and output conditions are set, and the parameters within the system associate large numbers of input conditions, in large sets of training examples. Coming to the specific case a particular input is set, and the system produces the trained correlating output as its answer.   This approach leaves the jury or decision-maker with no information about consistencies or differences in authorial style and essentially having to trust that the ipse dixit machine has come to a correct conclusion in any specific case. The jurors trust can only rest on the general case validation results rather than the specific case in question, giving a very real possibility of situations parallel to the PAN competitions, where there were hidden biases in the training set turned out to be crucial.

The solution to both the problem of bias in theory free development (the PAN competitions) and humans potentially reducing the accuracy of results (the mock jury findings), has to be a renewed focus on explanation.  Explainable AI (abbreviated to XAI) is a growing field, but it is rare to see advocates of AI taking on a basic user centred design approach.  User personas for a forensic evidence system (for authorship analysis) must include generalist police investigators or crime analysts, expert language analysts, and of course most importantly decision-makers such as jurors or judges in court.  The system must be designed to meet these individual user needs.  We need User Centred Explanation Design and as with User Centred System Design we need to develop methods to understand and document both commonalities and differences between these user groups and we must do this before we start to build any AI forensic science system.

Whilst these end-user analyses remain undone, we can adduce some general features of an explainable AI system for forensic evidence – the purpose of a good explanation will be to allow a human to calibrate trust and understanding in a system.  The end user in each group must be able to say 

  • globally – generally the system gets to the right answer in a way that I understand
  • in an individual case 
    • I can understand why and when  a system it’s likely to make good decisions and most importantly, when it might make mistakes.  Mistakes can include result with inflated or deflated likelihoods, for example. Where there is a useable explanation such  mistakes doesn’t corrode generalised trust in the system – “The system is wrong, but not nuts! I understand why it says that.”
    • I can the system must provides sufficient evidence of link between inputs and outputs.  This enables me to trust the chain of understanding and to make rational decisions, particularly against contrary independent evidence from other forensic sciences and types of evidence. 

This is may be high bar for explainable AI systems but there are already forms of XAI that begin to meet these criteria.  Until these techniques reach maturity, and until the explanation needs of specific end-users in forensic cases are properly understood and accounted for,  AI produced forensic evidence should not be admitted as evidence in court.

On the inadequacy of justice

Today I’m on holiday and today this story broke – Bernadette Walker: Scott Walker guilty of murder – BBC News. And this is a fuller version from the Peterborough Telegraph. I knew the verdict was coming, I hoped it was coming, and yet it’s been getting me down.

Once again a man has abused and then killed a young woman and once again that man is a member of the woman’s family.  But this case is unusual. 

In this case, the case of the murder of Bernadette Walker, this man who abused and killed his stepdaughter has been brought to justice.  Scott Walker has been convicted of Bernadette’s murder, in spite of the fact that Bernadette’s body has not been found.  Both Scott and Bernadette’s mother, Sarah Walker, have also been found guilty of perverting the course of justice, in part for sending messages from Bernadettte’s phone, which suggested that she was alive after Bernadette had been killed.   

This was the small part we played in this case. I worked with Márton Petykó to provide an expert report to the police. Our evidence concluded that the messages were inconsistent with Bernadette’s texting, and consistent with either or both of Scott and Sarah. (We couldn’t reliably separate their styles from one another). Sarah Walker admitted sending these messages before the trial. The jury today found that Scott was also guilty of perverting the course of justice with regard to sending these messages and they decided that Sarah must have known or believed that Bernadette was dead, which she had denied. As some of our evidence was agreed and some was less important to the overall case, the prosecutors decided there was no need for our evidence to be put before the jury, and so we didn’t give evidence in person.

This then is justice working.  This is what is meant to happen.  Individuals are held accountable and punished.  We made our limited contribution to the process.  Sentencing will occur in September – Scott will receive a mandatory life sentence for murder, and the judge will have to decide both on Scott’s minimum term in prison and on Sarah’s sentence for perverting the course of justice.  And as we played a small part in this case. I should feel good about this.  But I don’t. I feel depressed and weary.

Forensic linguistics is sometimes defined as a list of topics, or applications, or practices – but I prefer functional definitions.  My definition for forensic linguistics is that it is an attempt to improve the delivery of justice using the analysis of language.  This definition allows a broad focus, including improving legal and judicial processes, improving access to justice, and also  addressing investigative linguistic tasks like we did in this case.

In a small way in this case we have improved, or helped to ensure, the delivery of justice – so that’s good.  It’s also clearly inadequate.  

Criminologists sometimes write about five functions of sentencing, most often when discussing prison sentences. These five functions are punishment, prevention, rehabilitation, restoration and deterrence.

Scott and Sarah Walker, with their convictions and sentences are being punished and prevented from doing further harm whilst in prison, there is also some moderate hope that they will be rehabilitated, to prevent further crimes. 

Restoration though? Bernadette doesn’t get her life back, all those opportunities for friendships, and love and laughter. The opportunity for learning that there are good things in life.  These can’t be given back. That can’t be done. There is no sentence that can bring restorative justice for Bernadette. 

Deterrence seems like an easy target to hit but if we look around we see it really isn’t. Although there are many successful convictions, the violence continues.  Particularly violence against women.  This may be because even as increasing numbers of women are being abused by men (Women’s Aid stats suggest 1.6M domestic abuse cases a year in the UK and rising) most cases of abuse do not result in charge, let alone in conviction. In contrast and thankfully, most murder cases are solved – there’s a roughly 80% conviction rate in murder trials. I’ve not done the proper research today, but media reports suggest that there are currently about 2600 unsolved murders in the UK (some stretching back decades) and that the great majority of victims in these unsolved cases are also women.  

It seems to me that deterrence is failing because of a lack of recognition of the links between our misogynistic culture and the abuse and murder of women. The misogynistic culture we live and breathe, online and off, is all-pervasive and feeds the statistics of violence and abuse.  Rather than thinking about deterrence we need improvement of justice at this social level – today (in my weary and depressed mood) it seems that social justice in this sense is deteriorating. Improvement seems non-existent or at best imperceptible.  The scale of the required change is daunting.  I don’t know how you would measure such a change. Today at least I don’t know how you would target it, or who needs to do it, or even what insights and knowledge are required to make changes at a level and scale that would make any difference. I do know we need to figure this out and we need to do more.

I applaud all those who have achieved the delivery of justice in the Bernadette Walker case. Bringing these convictions was a difficult task and a task that was ultimately done well. For those who work day-in and day-out on such cases – I don’t know how you do it. I’ve probably been working on cases like this for too long, but I do so only occasionally and on days like today it exhausts me. My feeling today at least about the convictions in this case – the contribution to the delivery of justice that these convictions represent – is that both it is to be applauded and also that it is simply not enough.

[minor tweaks, updates and additional link 27th July]

Did AC12 hire a bent forensic linguist?

The big reveal in the season finale of Line of Duty was that it was a forensic linguist, who identified ‘H’ aka “the Fourth man” not as some criminal mastermind, but as the bumbling and over-promoted DI Ian Buckells.  The TV police drama has followed the incremental investigations of anti-corruption unit AC12 of the fictional Central Police.  Over the course of six series AC12 has established links between Organised Crime Groups and institutionalised police corruption, and in doing so it has traded on a tension between procedural hyper-realism and credulity-stretching flights of fancy.  The final message of the final episode appears to be that competence can become an ethics issue.  Buckells was no Moriarty, his corruption arose from inability, or lack of enthusiasm in doing his job well.  It was disappointing therefore that the culmination of AC12’s investigation relied on the evidence of a less than competent forensic linguist.

The linguistic plot-twist of the show rested on the spelling “definately” in the communications from the anonymous H. This misspelling had been flagged in an earlier series where we saw Superintendent Ted Hastings, use “definately” in an online chat, in which he assumed the linguistic identity of H.  H was discovered by searching though police files for this misspelling, and in the final interview scene a report is produced from a forensic linguist, that declares H is in fact DI Buckells. 

Each of these events relate to tasks which have been researched by forensic linguists.  Research in linguistic identity assumption shows that identifying lower-level linguistic features such as spelling patterns are necessary to successfully takeover someone’s identity in an online chat, but they are not enough.  Unless Hastings is also able to copy H’s linguistic turn taking and topic control, he is more likely to be discovered.  This is difficult but can be trained.  Our experiments with trainee undercover online officers showed that forensic linguistic training reduced detections by suspicious interactants from about 75% to 25%.  A further issue for Hastings is that he needs to avoid linguistic leakage of his own identity. He would be wise not to drop in the phrase “Mother of God” when chatting as ‘H’.

Using language style as a basis of an author search is also subject to intensive research.  One focus is to use writings by offenders in darkweb fora as a basis for searching for an individual’s writings on the open web.  Author search was used in this way as one of many techniques in the investigation that identified Matthew Falder, one of the UKs worst online sexual abusers. Another example is that the single-word greeting “Hiyas” contributed to the identification of paedophile, Shannon McCoole.  Searching written style in a fairly small collection of police files is significantly easier than searching all of the Internet.

It is the comparative authorship analysis task in Line of Duty , that is perhaps most known application of forensic linguistics, and in the real world this is a task that can and does produce evidence that is taken to court.  The dialogue of Buckells’ investigative interview in Line of Duty would have us believe such evidence can be very strong. Buckells is told that on the basis of the spelling “definately”, and a brief mention of some syntax features, that AC12s forensic linguist “concludes there is a 95% probability the messages detected on GGM-13 were written by you.” This is of itself a red flag suggesting that AC12s forensic linguist is either incompetent or perhaps, for reasons unknown, framing Buckells to take the hit.  

The tell that this is an inept or perhaps bent forensic linguist is the cited 95% probability.  This is not only bad forensic linguistics, it is also poor forensic science.  Contemporary research in forensic science suggests that identification evidence as should not be provided as a conclusion of match or non-match, even with a mitigating probability score.  Such an expression of the result is both  logically incorrect and usurps the role of the jury.  It is the job of the forensic scientist to express the weight of their evidence, and this is best achieved by taking a competing hypothesis approach.  In AC12s case we can easily create the two competing hypotheses with regard to the spelling “definately”.  These would be expressed as 

  1. “the probability of seeing “definately” given the hypothesis that Buckells wrote the ‘H’ messages; and,
  2.  the probability of seeing “definately” given the hypothesis that anyone else wrote the ‘H’ messages.

To calculate the probability for the first hypothesis you might trawl through all of Buckells’ and H’s writings to come up with a rate of occurrence per thousand words for each of them.  This gives some measure of similarity between Buckell’s and H’s writings in this regard.  But similarity is not enough.  To calculate the probability of the second hypothesis we need to research some base-rate information.  This is where things get tricky.  A quick google search gives me nearly 16,000,000 hits for the spelling “definately” and it is unlikely that all of these are references to Line of Duty.  The conclusion though is that this is not a rare spelling error and if this spelling is common then the similarity has little evidential weight.  If the spelling can be shown to be rare the evidence might then carry more weight. This weight of evidence can be best expressed as a ratio of the probabilities of the two hypotheses although studies show juries struggle to understand these Bayesian likelihood ratios when expressed mathematically.

It’s not enough however to only know population rate for the “definately” spelling.  A forensic linguist is above all a linguist and understands how language varies across communities.  Spellings which are globally rare can be locally common.  Research would be required to show the differing rates of occurrence of “definately” in the staff of the Central Police rather than in the general population. Spelling variants will also be more or less common across different genres of communication as lots of creative spellings are acceptable in online communications such as internet chat.  The most difficult and time-consuming part of providing a forensic authorship analysis is identifying and collecting the best comparison samples.  There are always compromised decisions in this respect.  Can I collect the language of police generally, or should I focus just on Central Police? Can I use the email database when the questioned messages are online chat? And so on. These are just some of the reasons why a lack of certainty can be a true mark of a forensic linguistic expertise. 

I would couch my report in terms of points of distinctiveness and consistency across the questioned samples, the known samples and the comparison corpus.  I would definately never use 95% certainty.  AC12’s forensic linguist however, appears to be uninterested in base rates, uninterested in the cross-genre nature of their analysis, and has expressed their result as a percentage certainty of match.  We have grounds to ask are they incompetent, or are they bent?

Tim Grant – Forensic Linguist

I am Professor of Forensic Linguistics and Director of the Aston Institute for Forensic Linguistics.

I am an academic practitioner and a practitioner academic in the field of forensic linguistics.

I am an experienced academic, skilled in teaching, productive in research and engaged in a range of administrative roles.  

I have attracted research income in excess of eight million pounds since I joined Aston University in 2007 and I led the development of the Centre for Forensic Linguistics (CFL) into the Aston Institute for Forensic Linguistics as the internationally leading research group in the discipline.  

I am also a practitioner academic – my casework has helped resolve numerous cases nationally and internationally.  These cases relate to the criminal law, issues of national security and civil matters:

  • I have contributed evidence in many criminal trials, for prosecution and defence, in cases involving terrorism, murder and sexual assault.
  • I have been involved cases where forensic linguistic evidence was critical in providing evidence for arrests where several hundred lives were at issue.
  • I was involved in the Yukos arbitration hearings which upheld a $50bn award against Russian Federation.  

Work such as this has drawn media attention and I am at ease engaging with press, television and radio.