AI Legal Slop and the Illusion of Legal Safety

A few months ago, the managing director of a successful Bahraini enterprise came to us late in a transaction. He was acquiring a controlling stake in a local competitor and, aiming for efficiency, had run the seller’s proposed share purchase agreement through a popular AI tool.

The output the AI generated looked remarkable. It produced a crisp, authoritative memo stating that the warranties were standard, the indemnity caps were within market range, and the completion mechanics were tidy. The AI’s verdict was clear: subject to a few minor typos, the agreement was balanced and ready to execute.

He was not coming to us for a full review; he just wanted a quick final blessing. He believed he was holding a shield. In reality, he was holding what is rapidly becoming the greatest new risk in corporate law: “AI slop.” It was a highly polished, deeply persuasive legal summary that was entirely unfit for purpose — and, frankly, dangerous.

The distance between a confident AI summary and actual legal safety is where the next decade of legal practice will be decided.

What the AI tool could not see — because it only reads paper, not people or markets — was a pattern of small signals surrounding the target. During our conversation, the client casually mentioned a few oddities: the target’s finance director had abruptly resigned two weeks prior without explanation. A footnote in the auditor’s recent management letter had a strange, evasive tone. Most notably, the seller’s principal, usually a rigid negotiator, had been unusually flexible on a payment schedule that no seller in Manama should bend on.

To a layman, these were minor commercial quirks. To the AI, they didn’t exist at all. But taken together, they are the exact pattern experienced counsel in Bahrain learns to read as contingent-liability risk — a liability that has not yet surfaced, sitting by design outside the AI-approved warranty package.

Because the AI was asked, “Is this agreement legally sound?” it answered “Yes.” It could not infer the question the client actually needed answered: given the behavioural red flags surrounding this deal, what is our real exposure, and how do we build a structural trapdoor to catch it?

The answer was not a blind renegotiation of warranties, which would have killed the deal and damaged relationships in a tight market. The answer was a carefully structured earn-out tranche tied to the very indicator the pattern was pointing at. It took a week to negotiate. Eleven months after signing, a liability of the exact kind the signals had pointed at surfaced. The earn-out absorbed the greater part of it without disruption. The relationships held. The transaction held.

Had the client signed based on the AI’s green light, he would have absorbed the entire loss.

This matter captures the shift we are watching across the profession. AI has not saved lawyers time. It has transmuted effort — moved it from one place to another — and in doing so, it has fundamentally changed what a law firm is actually selling. We have moved from an execution-heavy profession to a judgment-heavy one, and five consequences are becoming clear.

1. Legal work is not a deliverable

The more dangerous misconception about AI in legal practice is not that AI cannot produce a passable document. It can, and increasingly will. The misconception is that a legal document is what legal work actually is.

A memo, a contract, a pleading — these are artefacts. They are snapshots of a position at a moment in time. The work itself is a strategy, implemented through time: a continuous effort to navigate a shifting field of probabilities, counterparty behaviours, regulatory postures, and relational consequences, recalibrated over months or years as circumstances move. Even experienced lawyers — people who have spent decades inside a specific market — find this difficult to get right. It requires reading signals that have not yet surfaced, anticipating moves by counterparties whose incentives are not fully visible, and calibrating pressure to the tolerances of specific institutions in a specific jurisdiction.

A generic language model, trained on a vast but undifferentiated corpus of legal text, cannot run this kind of strategy. It can produce the artefact. It cannot read the landscape the artefact sits inside, nor the time dimension across which the matter will actually unfold. And the more confident the artefact looks, the more inviting it becomes to mistake it for the work.

The trap is that the output looks like the work. It is not.

2. Two distinct muscles: framing and discernment

High-level judgment is not a single skill. It is two, and they are trained differently.

Upstream — framing the question. What is the actual legal problem here? What are the boundaries of the exploration? AI has no access to the client’s history, their unspoken commercial objectives, or the political realities of their board. It cannot decide what question is worth answering. That remains a deeply human task.

Downstream — discernment. When a model returns five plausible legal strategies, which one is actually right for this client, this counterparty, this week? Discernment is the ability to filter options against a client’s unstated fears, their risk appetite, and the long arc of the relationship. AI produces the options. Only a seasoned lawyer can make the choice.

3. The tacit knowledge moat

We are told constantly that AI threatens “expertise.” But expertise is not monolithic. Codifiable expertise — standard drafting, basic research, first-pass document review — is eroding quickly, and that erosion is permanent.

Tacit expertise is a different category entirely, and it is insulated. The ability to read a room during a tense mediation. The instinct for when a regulator in Bahrain is signalling flexibility versus firmness. The judgment to know when to push and when to let a negotiation breathe. None of this lives in a dataset. None of it can be prompted.

The firms that thrive over the next decade will be the ones that stop competing with machines on processing power and double down on the unteachable.

4. The compounding cost of a small misjudgment

A legal matter rarely turns on a single decision. It turns on a chain of them — each one a judgment call, each one carrying its own sliver of risk. The opening position in a shareholder negotiation. The structure chosen for a new entity. The concession offered in a preliminary hearing.

This is where AI is most likely to go wrong — not at the level of the individual document, but across the arc of the matter. A model can produce a plausible opening position. It cannot see that the opening position, however reasonable on its face, quietly forecloses a settlement route the client will want twelve months later. Long-horizon strategic judgment is a function of accumulated experience inside a specific system. Catching a misstep before it happens is deep, local expertise doing its quiet work — and it is the part of lawyering that AI is furthest from replacing.

5. The illusion of coverage

There is a particular kind of legal danger emerging fast, and it may prove more costly than the risks most businesses currently worry about. It is the risk of acting on a legal position that looks well-constructed, reads persuasively, cites the right provisions, and is, nonetheless, wrong where it matters.

Unsolicited risk is at least visible; when a company knows it is exposed, it behaves carefully. The deeper danger is the opposite: a polished, AI-generated memo that creates the confident sense the risk has been handled, when in fact the analysis has missed a controlling statutory carve-out or misread the commercial intent of a clause.

This is the illusion of coverage — and it is uniquely destabilising because it changes behaviour. Armed with what feels like a defensible legal position, a business owner commits to a structure, signs the agreement, or absorbs the cost of a course of action they would never have taken had they understood how thin the ground beneath them actually was.

A significant number of businesses in this region will, we suspect, learn this lesson the hard way over the next few years. The hedge against it is not more AI. It is a trained legal eye that can tell the difference between an answer that is well-written and an answer that is actually right — and that knows which questions, in this jurisdiction, do not yet have the clean answer the model has just confidently produced.

The firm that remains

The technology will keep moving. What will not change is that a client in a genuinely difficult situation needs a lawyer who can sit across the table, understand what is actually at stake, and own a recommendation.

That is the work we are building Sahwan Law around — not competing with algorithms on output, but delivering what they cannot: the framing, the discernment, the long-horizon judgment, and the trained skepticism to tell a persuasive answer from a correct one.

This article is a general commentary on legal practice in Bahrain and the GCC and does not constitute legal advice. Application of these principles to any specific facts requires tailored counsel.