I have been living in two different worlds.
The first world ran from developing a real interest in science in my teens to my late 40s. It was a world before generative AI, in which if you wanted to understand something you had to hunt it down yourself: textbooks, journals, conference proceedings, monographs, commentaries. You learned to sit with conflicting studies, to check references, to notice the same names appearing in different literatures, to recognise when a popular idea rested on surprisingly thin empirical ice.
The second world is the one we all now inhabit, in which I can type a question into an AI system and receive, in seconds, a fluent synthesis that would once have taken me hours or days.
It is tempting to describe this second world as “revolutionary”. In some ways it is. But here is the uncomfortable truth, at least from my side of the consulting room: AI is revolutionary for me largely because of what happened in the first world. Thirty years of reading science, and of being wrong, confused, and occasionally persuaded, are what make AI useful now. Without that long apprenticeship, the same tool would be as likely to mislead as to assist.
What AI Is Actually Doing
To understand why prior knowledge matters, we need to be clear about what contemporary AI systems are and are not doing.
Large language models - the systems behind most current “AI assistants” - are trained on vast quantities of text. Their job is to predict the next word in a sequence, based on patterns they have learned from that text. They do not “understand” evidence, mechanisms, or legal standards. They do not know whether something is from a high-quality trial, an opinion piece, a marketing blog, or an unmoderated internet forum. They see text and learn statistical patterns in language.
That text includes:
- peer-reviewed research papers
- textbooks and professional guidance
- popular psychology and self-help material
- advocacy websites and blogs
- training materials of variable quality
- everyday social media discourse
When an AI system answers a question, it is not consulting a mental filing cabinet labelled “valid law” or “good science”. It is generating a plausible-sounding continuation of text based on what it has seen before.
If the dominant online narrative about, say, “attachment”, “trauma” or a fashionable neurodevelopmental label is conceptually muddled, scientifically thin, or clinically unsafe, that muddle will be faithfully reproduced. The system has no internal concept of “this construct is contested” or “this label is not recognised in ICD-11 or DSM-5”. It only knows that many people have written about it in a confident tone.
In short: AI is a mirror for our collective language, not an oracle for our collective knowledge.
What Thirty Years of Reading Actually Buys You
Why, then, does AI feel so powerful to someone who has spent decades immersed in scientific and professional literature?
It is not because the machine has suddenly become clever enough to do my job. It is because, long before I ever used AI, I spent years developing what philosophers would call epistemic virtues - habits of mind that regulate how you form and revise beliefs.
Over thirty years of reading science and working in clinical, educational and legal contexts, you accumulate, often without noticing, several kinds of tacit expertise:
1. A sense of evidential weight
You learn to distinguish between:
- a single underpowered study and a replicated body of work
- an attractive hypothesis and a demonstrated effect
- a case series and a properly controlled trial
- a narrative review and a systematic one
When an AI summarises “what the research says”, that background sense of weight lets you ask, “Is this actually a mature literature, or just a lot of people confidently repeating each other?”
2. Pattern recognition for how fields go wrong
You see fads come and go. You remember when particular labels were introduced, how they were over-extended, and what happened when they collided with courts, schools, and families. You watch promising constructs wither under better data. AI has none of that historical memory. It treats all well-phrased text as equally eligible for reproduction.
3. Conceptual hygiene
You become sensitised to category errors: using a family-systems concept as if it were an individual diagnosis; importing a clinical metaphor into a medico-legal setting; treating descriptive formulations as if they were established conditions. When AI blithely mixes these levels, you notice. Without that sensitivity, the blend can look impressively coherent.
4. Legal and regulatory context
Working as an expert witness teaches you that not every popular term is an acceptable term. Some fashionable constructs are simply not recognised in major classificatory systems; others lack agreed diagnostic criteria or adequate psychometrics. Over time you internalise the fact that language is not neutral: certain formulations carry consequences in court, in education, and in policy.
5. Experience of being wrong
Perhaps most importantly, you build up a feel for your own fallibility. You have followed literatures that later turned out to be overstated; you have had to update your views. That cultivated humility makes you much more cautious about treating any single AI output as definitive.
The result is that when I use AI, I do not experience it as “doing my thinking for me”. I experience it as a particularly fast, articulate junior assistant who is very good at finding patterns in text and very bad at knowing which patterns are safe to believe.
AI as Accelerator, Not Substitute
From that vantage point, AI genuinely is transformational.
In my own practice, systems of this kind can:
- surface relevant case law, guidance, and empirical work more quickly than manual searching
- draft a first pass at a complex document, which I then critique line by line
- help map out the terrain of debates in new areas, giving footholds for deeper reading
But crucially, AI multiplies the skills you already have; it does not conjure them into existence.
For someone who has not spent years reading research or working with real cases, the same system can be actively misleading. The fluency and confidence of the language create what might be called epistemic over-trust: the feeling that a well-phrased answer must be a well-founded one.
Across clinical and professional practice, I see the potential for:
- AI-generated reports that use clinical labels not recognised in ICD-11 or DSM-5
- confident statements about “attachment styles” or “trauma responses” that bear little relation to how these constructs are used in rigorous research
- sweeping generalisations about neurodevelopment based on tiny, biased samples
- policy recommendations that align with popular discourse but not with NICE guidance
To a reader without deep prior knowledge, these outputs can look polished, sensible, even impressive. To someone who has lived through the debates, they ring alarm bells.
Why This Matters for Law and Other Professions
From where I sit, this has obvious implications for law, but also for education, social care and health services.
There is nothing intrinsically improper about using AI to draft and prepare things. The pressures on time and resources in knowledge-based practice are real, and tools that can speed up routine work are understandably attractive.
The concern, from a psychological and scientific perspective, is that AI output can easily be mistaken for an impartial synthesis of the best available knowledge, when in fact it is a statistically plausible remix of whatever text the system has seen.
In areas such as child protection, neurodevelopmental disorders, mental capacity, and personal injury, important distinctions often hinge on:
- the difference between recognised diagnoses and unvalidated constructs
- the evidential status of particular assessment tools
- the nuances of developmental trajectories and comorbidity
- the boundaries between legal concepts (for example “best interests”) and clinical ones
AI systems do not know where those boundaries lie. They will freely import language from training materials, online CPD slides, advocacy sites, and popular psychology into contexts where courts and tribunals require a very different level of evidential discipline.
How This Article Was Written
Of course it was written with the assistance of an AI system.
That is not a contradiction of the argument; it is a worked example of it.
Several things were true before I opened an AI interface:
- I already knew, in outline, what I wanted to say.
- I had spent years thinking about the philosophy of science, the limitations of research, and the particular ways in which psychology can go wrong in legal and educational contexts.
- I had clear guardrails in mind: recognised diagnostic systems, NICE guidance, and the distinction between formulation and diagnosis.
- I was already uneasy about the way AI might reproduce popular but poorly evidenced constructs.
What AI contributed was not the core ideas but the shape of the text:
- organising arguments into a coherent structure
- offering alternative phrasings and transitions
- helping to tighten or expand sections for a legally literate audience
- serving as a sounding board for emphasis and framing
In other words, the machine provided linguistic fluency and speed. The content - the concerns, the distinctions, the insistence on certain guardrails - came from prior human knowledge and values.
Over time, because I have repeatedly corrected and constrained the system, its outputs have become better aligned with the way I think and work. It has “learnt from me” only in the limited sense that I have continually shaped its behaviour: rejecting certain formulations, reinforcing others, and specifying the standards I want it to observe. That is still not understanding on its part; it is a form of responsive pattern-matching tuned by a user who already has a sense of what good looks like.
If you removed that long apprenticeship - the years of reading science, the legal and clinical experience, the familiarity with what not to say - and gave the same AI system to someone without that background, you would get a very different result. It might be stylistically neat, but it would be far more likely to:
- uncritically repeat fashionable but unvalidated concepts
- blur the line between popular discourse and formal guidance
- overstate what the evidence can actually support
The fact that this article is, I hope, more careful than that is not evidence that AI can think. It is evidence that AI can be very useful in the hands of someone who has already done a great deal of thinking.
The Operator, Not the Tool, is the Safety System
A helpful analogy is aviation. Modern aircraft are full of automation. They can, in principle, take off, fly and land with very little human input. Yet we do not respond by putting untrained passengers in the cockpit. If anything, the competence required of pilots has increased: they must understand both the underlying physics and the behaviour of the automation, including its failure modes.
In a similar way, AI does not reduce the need for human expertise in law, psychology, education or medicine. It changes what expertise needs to include.
In my own field, safe use of AI is closely tied to:
- robust domain knowledge (clinical science, law, education)
- familiarity with relevant regulatory frameworks (NICE, ICD-11, DSM-5-TR, statutory guidance)
- the ability to evaluate sources and spot unvalidated constructs
- an understanding that AI models do not know what they do not know
- a willingness to treat AI output as a draft to be interrogated, not as an answer to be accepted
Other professions will have their own equivalents. The common thread is that the safety system is supplied by the operator, not by the tool.
A Different Kind of AI Literacy
Much of the current conversation about “AI literacy” focuses on prompt techniques and basic awareness of bias and hallucination. Those are useful starting points, but they are not enough.
For professionals working in law, education, social care and health, the crucial literacy is deeper and older. It is the capacity to ask, whenever an AI system offers an answer:
- What is the underlying evidential base here?
- Which constructs are being used, and are they recognised and valid?
- How would I know if this was wrong?
- What are the legal and ethical implications if I treat this as true?
- Does this align with established guidance, or merely with popular discourse?
These are not questions that AI can answer for us. They are questions we bring to AI, from a background of training, reading, supervision, and practice that long predates the technology.
If I had to compress the point into a single sentence, it would be this:
AI is most powerful in the hands of people who would already be dangerous without it.
“Dangerous” here in the positive sense: capable, critical, and accountable. The kind of professionals who already know that fluent language and sound knowledge are not the same thing, and who are prepared to do the hard work of telling the difference - with or without a machine in the room.