Automation bias in legal work: who decides?
AI now drafts documents, summaries and entire court filings for law firms and legal departments in a few minutes. Responsibility and liability for the result stay with the person who approves it and files it. In October 2026, two German courts, the senior judiciary of England and Wales and the Law Society of England & Wales each made that point in their own way.
The open question is whether an approval is still an independent judgment or just confirms what the machine suggested. A study from the end of October shows how quietly that can shift; researchers call it automation bias. The article closes with a decision inventory that helps decide, per process, where a person has to judge (not legal advice, but a best-practice approach).
Several sources were published in German. Quotes from them are our own translations.
Key facts (as of November 2026)
- Anthropic study, 27 October 2026: 50 professionals followed an AI agent’s randomly assigned lean on analytical decisions in 86% of cases. Only 15% later mentioned the AI in their answers (heise, 30 October 2026).
- German Federal Labour Court and Bavarian Constitutional Court, September 2026: in both cases, filings cited decisions that do not exist. The Federal Labour Court took this as a sign that the lawyer had not checked the brief himself and dismissed the complaint as inadmissible.
- Law Society of England & Wales, October 2026: professional responsibility cannot be transferred to an AI tool. The statement is not binding and applies to England and Wales only.
- German Federal Bar (BRAK), October 2026: AI output must be checked in full and on substance; spot checks are not enough. The duty of care rises with the degree of automation.
- In practice: an approval at the very end does the least. It works better where the substantive decision is made, with its basis visible and recorded automatically.
Responsibility for AI drafts: recent decisions and context
Most signatures are electronic now, and German filings go through the lawyers’ electronic mailbox, the beA. What matters is who stands behind the content, and that remains a person, even when a machine wrote the text.
Two German court decisions from September show what happens when nobody has really checked the content. On 30 September 2026 (2 AZN 339/26), the Federal Labour Court (Bundesarbeitsgericht) dismissed a complaint against the refusal of leave to appeal as inadmissible: four cited decisions did not exist and three more were reproduced with incorrect content (LTO). The court took this as a sign that the lawyer had not checked the brief himself, so the representation by a lawyer entitled to plead, mandatory before the Federal Labour Court, was missing. On 23 September 2026 (Vf. 9-VI-25), the Bavarian Constitutional Court imposed an abuse fee of 3,000 euros on a constitutional complaint built on invented decisions, in the court’s view “probably with the help of so-called hallucinating AI” (beck-aktuell). Which checks belong before a filing goes out is covered in our article on reviewing AI-drafted court filings. This one is about the question before that: whether the person approving really judges for themselves, and who is liable when it goes wrong.
In October, the Law Society of England & Wales reminded its members in a statement that ‘AI is not a legal entity’ and that solicitors therefore ‘cannot transfer professional responsibility’ to an AI tool. The statement is not binding and applies to England and Wales. In substance, it would surprise few lawyers in Germany either.
A WDR report of 30 October on the social courts in North Rhine-Westphalia shows how concrete this gets. A specialist social-law lawyer describes a request: the objection had already been drafted, and the client “only wanted me to put it on my letterhead and send it off as it was.” The letterhead was what mattered, because the person behind it vouches for the content. The pressure is growing: according to WDR, urgent applications to the social courts in North Rhine-Westphalia rose by 157 percent in the first quarter of 2026 compared with the first quarter of 2025, from around 1,400 to 3,600.
What this means for liability
A lawyer who adopts an AI draft makes it their own work. If it contains an error, that error counts as their own – towards the court as much as towards the client. Berlin’s Higher Regional Court (Kammergericht) said so in November 2025 (17 WF 144/25): a lawyer is obliged to check AI-inserted citations “both under the mandate and as an officer of the administration of justice.”
The German Federal Bar (Bundesrechtsanwaltskammer, BRAK) spelled this out in its updated guidance on the use of AI in October 2026. It requires the lawyer’s own review and final check of AI output, in full and including the cited sources: “spot checks are not enough.” And: “the duty of care in dealing with AI rises with the degree of automation.” In the BRAK’s view, failing to check can also breach the duty of conscientious practice under section 43 of the Federal Lawyers’ Act (BRAO). According to LTO, the Federal Labour Court itself referred to the BRAK’s AI guidance when it set out the duty to check in case 2 AZN 339/26.
One point matters especially for approval: in the context of labelling under Article 50 of the EU AI Act, the BRAK states that “a purely formal or linguistic check” is not enough. An approval that only checks whether the text reads well is therefore not a review in the BRAK’s sense.
The Federal Labour Court case shows that this can get expensive. If an appeal is treated as inadmissible because of an unchecked brief, the client has lost that instance. The question of whether the client suffered a loss as a result follows quickly. That is the classic starting point for a professional negligence claim, and the tool used changes nothing about it.
Three practical consequences follow:
- The vendor is no liability shield. A firm has room to negotiate liability in its contract with its AI vendor. The lawyer’s liability towards the client, however, remains entirely unaffected.
- Documentation is evidence. In a dispute, only what was recorded helps: who checked what, when and on what basis. “Someone always looks it over” cannot be proven.
- Talk to your insurer. The BRAK points to a risk many overlook: if automation goes so far that the work is no longer legal work by a lawyer, the professional indemnity insurer may refuse cover in a claim for lack of lawyerly activity. Lawyers should clarify this with their insurer before an actual claim arises.
For in-house legal departments the question looks different, because the company itself is the client. There too, the person who approves must be able to show internally what the decision was based on.
Automation bias: why an approval is not automatically a judgment of your own
So far, the debate about AI in legal work has mostly been about whether anyone checks. Who reviews the draft, who verifies the citations? The principle behind it is called human in the loop: at a defined point in an automated process, a person checks or decides. It assumes that the reviewer judges independently. A study by Anthropic, published on 27 October 2026, calls exactly that into question.
In it, 50 data-analysis practitioners worked with an AI agent on a task involving open questions of judgment. Which way the agent leaned on each decision had been set at random beforehand, and participants were not told. heise summarises the result: in 86 percent of the analytical decisions the participants went with the model’s view, which in this experiment was not the result of reasoning but a roll of the dice. Without any influence from the AI, 50 percent would have been expected. And: only 15 percent of participants mentioned the AI agent when explaining their decision.
The study has clear limits. The participants were not lawyers and were new to the task’s subject area, the group was small, it covered a single task with a single model, and the AI’s preferences were deliberately amplified for the experiment. The study is a preprint without external peer review; its design was pre-registered. The second number is more striking than the adoption rate: hardly anyone noticed that the AI had any influence at all. At the same time, 37 of 48 participants repeated the agent’s arguments in their own answers.
Professor Rolf Schwartmann explained the reason to beck-aktuell on 20 October: “People tend to find suggestions from machines especially plausible.” Research calls this automation bias. Schwartmann’s answer is a sequence of human, machine, human. Thinking first yourself makes it less likely that “the human is reduced to a lifeguard on the poolside who just nods things through at the end.”
Emily Bull, a lawyer in the legal team of the energy company SSE, describes the long-term side in an interview with Artificial Lawyer on 29 October: “declining verification skills concerns me particularly because it could happen quite quietly.” People who mostly confirm gradually lose the habit of checking and hardly notice.
Sir Geoffrey Vos, Master of the Rolls in England and Wales, named the underlying dilemma in his speech at Legal Geek on 14 October. He spoke of “the stark fact that AI only saves the time of human lawyers and judges if those lawyers and judges actually accept what it says”. Reading everything again saves nothing. Reading nothing leaves you carrying the risk alone. The way out lies in the process: the part a person really has to judge should stay small and visible.
Where the approval should sit
Angelo Aratari put it this way in LexBeat #1, the magazine of the German Legal Tech Summit, on 14 October 2026: “So control hasn’t disappeared. It has just moved to other places.” It used to be built into the work. Anyone drafting a brief from scratch checked citations, figures and the line of argument along the way. When a machine writes the text, that silent check disappears, and someone has to place it deliberately somewhere in the process.
Out of habit it often ends up at the very end, as a last look at a finished document. That is where it costs most. Someone who sees only the result has to reconstruct in their head which inputs went in, which sources the AI used and where it made an assumption. That either takes so long that the time saving is gone, or it turns into skimming. A finished, well-written text also feeds automation bias, because it sounds right.
Approval does more where the substantive decision is made. André Moser, President of the Administrative Court of Zurich, described in the NZZ on 15 October what good lawyers do: “Good lawyers also tell their clients when something isn’t worth it and that they should stick to the essential arguments.” And: “AI doesn’t have that filter.” Sometimes the most important check is therefore the first one: whether a step is worth taking at all, before anything is drafted.
In legal departments, too, AI is moving to the point where requests come in. Vendors want it to work there with material that beck-aktuell (15 October) describes as the legal department’s stored work instructions and guidelines. Such a front door is only as good as the rules that are actually written down, and the point at which a person approves the results.
The decision inventory: one process, one table
A ban helps as little as blanket suspicion of every AI output. It makes more sense to sort the decisions in a process deliberately, once, before the division of labour with AI settles in by itself.
For a first sort, use the question Emily Bull applies to every legal task: “can we Eliminate the task, Enable someone else to do it, Engineer a better way of doing it, or does it genuinely need a lawyer’s Expertise?” Her team deliberately started small: “We didn’t start with a huge mapping exercise.”
For the steps that remain, answer five questions. The right-hand column gives an invented example: the first assessment of whether a contract from the business needs legal review.
| Question | What matters | Example: first assessment of a contract request |
|---|---|---|
| 1. Rule or judgment? | Deadlines, responsibilities, thresholds and mandatory fields can be set once as a written rule. Genuine questions of judgment need a person. | Contract value, contract type, missing mandatory data: rule. Whether a deviating liability clause is acceptable: judgment. |
| 2. Who decides, with what authority? | A role, not just a name. May the person only confirm, or also change, reject and escalate? | In-house counsel in the legal department. She can change the result and escalate to the head of legal. |
| 3. What does the approver see? | The inputs, the path through the branches, the sources behind an AI answer and the points where AI was involved. | The form data, which rule applied and which passage of the contract supports the AI summary. |
| 4. Where does AI only suggest, and who adopts it deliberately? | Adopting a suggestion is a decision of its own. The sequence human, machine, human helps. | The lawyer reads the clause herself first, then the suggested risk rating, and records why she agrees or deviates. |
| 5. Where is it recorded who decided what? | Automatically with every run, not by hand afterwards: who, when, about what, on what basis, with what outcome. | One report per request. Changed and rejected suggestions can be analysed later. |
This often shows that far more steps follow fixed rules than expected. Schwartmann gives air passenger rights as an example where generative AI is often not needed at all, because expert systems with if-then rules are enough. The Schlichtungsstelle Reise & Verkehr (SRUV), the German arbitration board for travel and transport that also handles air passenger cases, shows what that looks like. It handled more than 45,000 cases in 2025. Legal bots in e! classify each case by fixed rules and use AI only where the complaint and the company’s response have to be read. Every classification comes with a documented chain of reasoning, and people decide at the steps that require evaluation. Preparing a case now takes about one minute instead of five to seven (read the case study). Every rule that has been checked once shrinks the area a person has to approve on substance, and the genuine judgment calls get more attention.
One process that comes up often and regularly calls for judgment is enough for a first inventory. If the result holds up, add the next one.
The inventory works with any tool, even on paper. In e!, these decisions sit in the decision tree: the experts place the approval step themselves where it belongs, AI works only where someone has put it, and every run is documented. How to write a process down first and decide for each step between logic, AI and a person is covered in Legal AI adoption: map the process before you pick the tool.
Our Legal Use Case Canvas works as a template for a first inventory: one page, eight fields, with four filled-in examples from legal departments and law firms.
Frequently Asked Questions
What is automation bias, and why does it matter for approvals?
Automation bias is the tendency to trust a technical system’s suggestions more than they deserve. It is strong at the approval stage because the reviewer sees a finished, well-written result and often not the inputs and sources behind it. What helps is a sequence in which the person assesses first, and an approval step that shows the basis of the result.
Who is liable if an AI draft is wrong?
Not the AI: it is not a legal entity. A lawyer who approves or files an AI draft makes it their own work; Berlin’s Kammergericht (17 WF 144/25) and the BRAK’s AI guidance of October 2026 both require a full review on substance. Liability in a specific case depends on the applicable law and the circumstances.
Which steps in a legal process don’t need AI?
Anything decided or calculated by fixed rules: mandatory fields, responsibilities, deadlines, thresholds, calculations and approved standard clauses. Such steps produce the same result for the same inputs and can be checked once rather than in every case. AI is useful where unstructured language has to be understood or a draft written, and that is where approval should sit.
How do you document an approval so it can be traced later?
Record who decided, when, about what and on what basis: the inputs, the path through the branches, the sources behind an AI answer and the decision itself (approval, change or rejection). Documentation created automatically with every run is the most reliable. It lets you show later that a review took place, and it shows where results are frequently corrected.
Who should decide where approval sits in a process?
The experts who own the process and stand behind the result: partners in a law firm or the head of the legal department, together with the colleagues who do the work every day. IT, a vendor or a legal engineer can help with implementation, but where a person has to judge remains a professional question.
This article describes good practice for using AI in law firms and legal departments. It is not legal advice and does not replace a legal review of a specific case.
Ready to automate your legal workflows?
Discover how e! can transform your legal operations with no-code automation.