GenAI, the Verification-Value Paradox – a Critique
There has been much chatter on LinkedIn about a new academic article about the Verification-Value Paradox (of GenAI use by lawyers).
The article claims that it is doubtful that GenAI delivers value to lawyers because any efficiency gains are erased by the time spent verifying its output; a framing that the author calls the “verification-value paradox.”
The “paradox” is:
More AI = more verification = less value.
The author admits that this paper is not based on fresh, robust empirical evidence, as he waffles back and forth in much of his discussion of the paradox; essentially stating that GenAI is sometimes helpful and sometimes not.
From the standpoint of a legal practitioner, this is the equivalent of saying walking across the road “can” be dangerous.
The decision to implement GenAI in a legal practice is to attain gains over elements of different files – not to say, “Well, if I can’t use it for legal research, then it is of absolutely no use to me whatsoever.”
The author’s ultimate, ordinary and unhelpful conclusion is, “for lawyers to be cautious, critical and/or reticent as to AI use in legal practice.”
This is precisely what law societies, bar associations and courts have been saying for the last 2 years. So, in other news, the author is telling lawyers that the sky is blue.
Ironically, the paper does little to support the proposition that GenAI is not useful for lawyers; the evidence is selectively presented, and the conclusion is not supported by practitioner studies or time-motion analysis.
Be that as it may, there are 5 important points that I wish the author had dealt with as they are far more interesting and relevant to practicing lawyers.
1. Verification isn’t a paradox; it’s the lawyer’s job
Lawyers verify everything: junior memos, discovery summaries, expert reports, and even partner-drafted factums. This isn’t burdensome; it’s what competent practice looks like.
AI doesn’t create a new obligation to verify. It shifts where verification happens and how much time we spend on low-value drafting versus high-value analysis. That’s workflow design, not a paradox.
The assumption that verification cancels out gains ignores basic realities observed in firms using these tools (RSG Study). A decent first draft (say 60% quality) saves meaningful time. Instead of starting from a blank page, the lawyer starts from a structured, reasonably coherent draft and moves immediately into:
(i) refinement,
(ii) strategy,
(iii) judgment, and
(iv) nuance.
GenAI is not designed to replace any of the above as these items are the core values that lawyers bring to the table.
Editing is faster, deeper, and more analytical than authoring every word from scratch – even with verification.
This framing – author to editor – is the much more interesting and relevant shift in the profession, but the article ignores it entirely.
In fact, the author spends a mere paragraph in his 28-page paper on “transactional and in-house lawyers who never engage with the court” and simply says the verification costs of GenAI use by these lawyers is also very high. Empirical practitioner studies and time-motion analysis are needed to support his assertion, but he has none.
2. The uncomfortable truth: many litigators don’t read the cases upon which they rely
The author suggests that because the need for accuracy is so high in litigation, the verification costs are therefore too high to use GenAI for legal research. However, he does not provide any data to support his view that verifying cases pulled by GenAI is slower than cases pulled and verified by other means.
He fails to mention that GenAI hasn’t created a verification crisis for legal research, it has revealed one.
For decades, litigation practice has tolerated:
(i) template-based arguments with inherited citations,
(ii) reliance on headnotes instead of underlying reasons, and
(ii) recycled factums passed from junior to junior.
GenAI didn’t cause this prior behaviour, but it is now shining a very bright light upon it.
Instead of blaming the technology and suggesting that it is causing problems, we should thank it for surfacing real problems in the profession.
3. It is absurd to use general-purpose LLMs for legal research, then blame Gen AI when it goes wrong
This may be the weakest assumption in the article: that using a general purpose LLM, such as ChatGPT or Gemini for legal research is a meaningful test of GenAI’s value in legal practice.
It’s not.
It’s negligence.
General-purpose LLMs are not legal research engines.
That’s like using a butter knife as a scalpel and concluding “surgery is unreliable.”
Wrong tool, wrong workflow, wrong conclusion.
The author cites only one South African Legal Research GenAI start-up (Legal Genius) as being his proof that legal-specific GenAI is unreliable; the bigger players in legal research including the author’s part-time employer, Thomson Reuters, are probably rolling their eyes at his example. Moreover, the use of only one example seems to be a leap unworthy of an academic paper, especially since as of September 2025, Legal Genius claims to have eliminated hallucinations using RAG.
4. The author assumes humans are more accurate than technology. They aren’t.
A silent premise in the article is that humans are inherently more reliable than machines.
Humans hallucinate too, it’s just that when humans, misstate judgments, rely on summaries instead of full judgments, forget to update cases, and simply get sloppy under fatigue or time pressure, we call it “human error,” “an oversight,” or “carelessness.”
Suggesting that GenAI’s fallibility somehow makes it unfit for legal work ignores the very real fallibility of the humans using it.
5. The article treats GenAI as if the technology is frozen
The technology in this area is moving rapidly. But the article leans on now-dated hallucination studies and ignores current tool design such as:
(i) retrieval-augmented generation,
(ii) curated legal information,
(iii) citation validation,
(iv) hallucination suppression, and
(v) firm-level governance.
It’s like assessing today’s satellite GPS accuracy using the performance of the very first civilian receivers from the 1990s. You might as well warn drivers that “GPS can be off by 300 metres.” True then, but absurd now.
This article will likely only be formally published in 2026, making it even more out-dated.
If anything, this article highlights the danger of academic journals publishing in a rapidly changing area; especially if the article does not have solid empirical data and it does not provide a solid value proposition to the profession.
Conclusion
There is no “verification-value paradox.”
The article leans on an inflated belief in human infallibility, the rehashing of well-known critiques and long-standing sloppy legal habits, then wraps it all in a catchy phrase built on outdated assumptions and the misuse of the wrong tools.
I must disagree, using your numbering and headings:
1. Verification isn’t a paradox; it’s the lawyer’s job
But it becomes a paradox when the verification is of hallucinated output.
The author’s concentration is on research and not, as you say, transactional work. That is where the main problems arise and where the verification burden is a real problem.
With regard to starting with a blank page (which lawyers seldom do) see: https://thetimeblawg.com/2024/09/01/blank-page-law/
2. The uncomfortable truth: many litigators don’t read the cases upon which they rely
Presumably then it is those litigators that are relying on GenAI!
You say “However, he does not provide any data to support his view that verifying cases pulled by GenAI is slower than cases pulled and verified by other means.” He doesn’t have to. It is obvious. If you have a made up case, with made up citations, that does not exist, that you have to find, you are never going to find it. It could take you quite a while to ascertain that. Case citations in legal textbooks, case citators and existing case law will not present that problem.
The litigation practice you refer to is thankfully not one I am familiar with. I’ve been practising for 3+ decades and have never seen this. Maybe things are different in North America from the UK?
If GenAI has surfaced real problems in the profession that does not obviate the GenAI verification burden. It also clearly shines a light on that.
3. It is absurd to use general-purpose LLMs for legal research, then blame Gen AI when it goes wrong
It may be absurd but many lawyers are, unfortunately, doing it. New cases of this are arising by the day.
Accepting that it is negligence to use them accepts that the verification burden is too high to use them. The author was, in my view, generous with his view of cancelling out the benefits. In my opinion, the dangers of GenAI in legal research far outweigh any perceived benefits.
However, the author is looking at what is actually going on in practice. You cannot ignore that.
The author states “Studies have repeatedly found ‘hallucinations’ even in machine learning tools built for the legal context. Even where leading legal research companies like Westlaw and Lexis have built AI into their search functions, they remain unreliable.”
4. The author assumes humans are more accurate than technology. They aren’t.
I can’t agree that humans hallucinate like GenAI. A lawyer (pre GenAI) would find case law in the way trained to do so using established legal research methods. That resulted in them locating and reading real case law. If they use GenAI, without proper verification, they are relying on nonsense that they would never otherwise have found or have been bothered by.
You state “Suggesting that GenAI’s fallibility somehow makes it unfit for legal work ignores the very real fallibility of the humans using it.” However, you contradict yourself in that you have previously stated that it is unfit for legal research. On who is to blame see: https://thetimeblawg.com/2024/01/14/lawyers-or-machines-who-do-you-blame-for-genai-hallucinations/
5. The article treats GenAI as if the technology is frozen
Hallucinations are still a big problem. It has been suggested that they are getting worse, not better, with new models: https://www.newscientist.com/article/2479545-ai-hallucinations-are-getting-worse-and-theyre-here-to-stay/
RAG does not eliminate hallucinations and the verification burden remains. And obviously you need to ensure the knowledge base you are using contains the information that you need. The verification burden is as much about what will be omitted by GenAI than included in its output.
Conclusion
There is a “verification-value paradox” in certain situations. Not all. But many. It is dangerous to suggest otherwise. There has to be a realisation of what lawyers, on the whole, are actually doing. That is a real assumption and based on the tools that they are actually using.