What is human in the loop in legal AI?

Human in the loop (HITL) is a workflow in which AI output has no effect until a named, competent person reviews it and approves, changes or rejects it. In a law firm, the AI drafts, extracts or flags, and a lawyer checks the output against the sources and the client's facts before anything is sent, filed or applied.

What is the difference between human in the loop and human on the loop?

Timing. In the loop, a person approves each output before it takes effect. On the loop, the system acts and a person monitors and can intervene afterwards. For client advice, filings and contract positions, professional rules expect review before the output has effect; on-the-loop monitoring can suit low-risk internal tasks such as tagging or routing.

What does ABA Formal Opinion 512 require of lawyers using AI?

Issued on 29 July 2024, it applies the Model Rules to generative AI: competence under Rule 1.1, including an appropriate degree of independent review of output; client informed consent under Rule 1.6 before client information goes into a self-learning tool; supervisory policies and training under Rules 5.1 and 5.3; and fees under Rule 1.5 that reflect time spent, not time saved.

Does Article 14 of the EU AI Act apply to law firms?

Article 14 applies to high-risk AI systems. Most law-firm tools for drafting, research and contract review are not high-risk under Annex III, but tools that assist judicial authorities or migration decisions can be. Since Regulation (EU) 2026/1744, the Annex III obligations apply from 2 December 2027. Article 4 AI literacy duties have applied to deployers since 2 February 2025.

What does the SRA expect of solicitors using AI?

The SRA applies its Code of Conduct: solicitors remain accountable for work carried out through others and must supervise it (paragraph 3.5) and keep client affairs confidential (paragraph 6.3). In R (Ayinde) v Haringey [2025] EWHC 1383 (Admin), the Divisional Court added that firm leaders must put practical measures in place so AI research is checked against authoritative sources.

How many court cases involve AI hallucinations?

Damien Charlotin's AI Hallucination Cases database listed 2,202 court decisions involving AI-fabricated or misrepresented material as of its 19 October 2026 update. Our hallucination audit at /blog/ai-legal-hallucination-audit breaks the cases down by court, country and sanction.

Legal AI s člověkem v procesu: co vyžadují ABA 512, AI Act EU a SRA

In short: human in the loop means a qualified lawyer reviews and approves AI output before it reaches a client, court or counterparty. ABA Formal Opinion 512 (29 July 2024), Article 14 of the EU AI Act and the SRA Code of Conduct put that duty on people. Charlotin's database lists 2,202 court decisions involving AI-fabricated material (19 October 2026).

Definition. Human in the loop (HITL): a workflow in which AI output has no effect until a named, competent person has reviewed it and approved, changed or rejected it. Human on the loop: the AI acts and a person monitors and can step in. Human out of the loop: the AI acts with no human review.

Key facts

ABA Formal Opinion 512 on generative AI tools, issued 29 July 2024, applies Model Rules 1.1 (competence), 1.6 (confidentiality), 5.1 and 5.3 (supervision) and others, and requires an appropriate degree of independent review of AI output (ABA, 29 Jul 2024).
Article 14 of the EU AI Act (Regulation (EU) 2024/1689) requires high-risk AI systems to be designed so that people can understand, monitor, correctly interpret, override and stop them, and to stay aware of automation bias (EUR-Lex text, read 20 Oct 2026).
Under Regulation (EU) 2026/1744, the high-risk obligations for Annex III systems, Article 14 included, now apply from 2 December 2027 (Official Journal, 24 Jul 2026).
The SRA Code of Conduct for Solicitors keeps solicitors accountable for work done through others (paragraph 3.5) and bound by confidentiality (paragraph 6.3); in R (Ayinde) v Haringey [2025] EWHC 1383 (Admin), the Divisional Court said firm leaders must take practical steps so everyone checks AI research against authoritative sources (Judiciary of England and Wales, 6 Jun 2025).
Damien Charlotin's AI Hallucination Cases database lists 2,202 court decisions involving AI-fabricated or misrepresented material (database, last updated 19 Oct 2026).

What does human in the loop mean in legal AI?

The term comes from control engineering, where a person sits inside the decision cycle and the system waits for their input. In a law firm it means three things in practice. The AI drafts, extracts or flags. A named lawyer reviews that output against the sources and the client's facts. Nothing leaves the firm, and nothing is filed, until that lawyer approves it.

The difference between in the loop and on the loop is timing. If a lawyer reviews a draft before it is sent, the lawyer controls the outcome. If the system sends it and the lawyer reads a log afterwards, any error has already reached the client. For client advice, filings and contract positions, professional rules expect review before the output has effect. On-the-loop monitoring can be reasonable for low-risk internal tasks such as tagging documents or routing intake.

What does ABA Formal Opinion 512 require?

The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on 29 July 2024. It applies existing Model Rules to generative AI:

Rule 1.1 (competence): lawyers need a reasonable understanding of what the tool can and cannot do, and must review its output to a degree that fits the task. Relying on output without an appropriate degree of independent verification can breach the duty.
Rule 1.6 (confidentiality): before putting client information into a self-learning tool, which may reuse inputs, the lawyer needs the client's informed consent, and boilerplate in an engagement letter is not enough.
Rules 5.1 and 5.3 (supervision): partners and managers must set clear policies on permitted AI use, train lawyers and staff, and supervise the work, including vendors and their tools.
Rule 1.5 (fees): lawyers can bill for the time they spend, including reviewing AI output, but not for time saved, and not for learning a general-purpose tool unless the client asked for that tool.

Several state bars, including California, Florida and New York, have issued their own guidance with similar review duties. Our bar-rules tracker lists them by country and state.

What does Article 14 of the EU AI Act say about human oversight?

Article 14 of Regulation (EU) 2024/1689 applies to high-risk AI systems. It requires providers to design them so they can be "effectively overseen by natural persons" and requires measures proportionate to the risk, the system's autonomy and its context. The people assigned to oversight must be able to:

understand the system's capacities and limitations and monitor its operation;
remain aware of automation bias, the tendency to over-rely on output;
interpret the output correctly;
decide not to use the system, or disregard, override or reverse its output;
interrupt it through a stop procedure that leaves it in a safe state.

Most legal AI used by law firms (drafting, research, contract review) is not high-risk under Annex III. Annex III does cover AI intended to assist a judicial authority in researching and interpreting facts and law, and AI used in migration and asylum decisions, so court-facing and public-sector tools can fall inside it. Separately, Article 4 has required providers and deployers to ensure AI literacy among staff since 2 February 2025, whatever the risk tier. The 2026 amendments moved the Annex III high-risk dates to 2 December 2027; the detail is in our EU AI Act amendments explainer.

Even where Article 14 does not bind a firm directly, its list is a useful checklist for designing review: can the reviewing lawyer understand, interpret, override and stop the tool?

What do the SRA and the courts in England and Wales expect?

The SRA applies its existing Code of Conduct for Solicitors to AI use. Paragraph 3.5 keeps a solicitor accountable for work carried out through others and requires effective supervision; paragraph 6.3 requires client affairs to be kept confidential. The SRA's Risk Outlook report on AI in the legal market (November 2023) says firms stay responsible for the services they deliver with AI. When it authorised Garfield.Law, an AI-driven firm, in May 2025, the SRA said named solicitors remain accountable for the firm's outputs.

The courts have been more direct. In R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin), handed down on 6 June 2025, the Divisional Court (Dame Victoria Sharp P and Johnson J) dealt with fictitious authorities in submissions. It said lawyers who use AI for research must check it against authoritative sources, that those with leadership responsibilities in firms and chambers must put practical measures in place, and that placing false material before a court can lead to referral to the regulator or contempt proceedings.

Vyzkoušejte HAQQ AI zdarma

Zažijte sepisování a rešerše s podporou AI

Which AI errors does human review catch?

These are the errors we see most often when lawyers test AI on real matters, and each needs someone who knows the law and the file:

Fabricated or misdescribed authority. A case, section or quotation that does not exist, or a real case cited for a point it does not support. Charlotin's database counts 2,202 court decisions on this. Our hallucination audit breaks the cases down by court and sanction.
Wrong jurisdiction. Common-law reasoning applied to a civil-law contract, or GDPR concepts imported into a Saudi data processing agreement governed by the Saudi Personal Data Protection Law.
Missed interactions. A limitation of liability clause that reads as standard until a carve-out three sections later changes it, or a side letter the AI never saw.
Confidentiality exposure. Client documents sent to a consumer tool whose terms allow it to keep or reuse inputs. In United States v. Heppner (S.D.N.Y., February 2026), documents a defendant generated with consumer Claude were held protected by neither privilege nor work product.
Overconfident output. A clause marked "standard" with no reasoning attached, which invites the reviewer to skip it. Article 14 names this automation bias.

There is an older lesson outside law firms too. In Moffatt v. Air Canada, 2024 BCCRT 149, the British Columbia Civil Resolution Tribunal held the airline liable for a refund policy its website chatbot invented, and rejected the argument that the chatbot was responsible for its own statements. The organisation that deploys the AI answers for it.

How do you design human review that works?

Put the gate in the workflow. Drafts, risk flags and redlines should need an approval step before they can be sent, filed or applied. An approval button in the tool is harder to skip than a policy people must remember.
Show the reasoning and the sources. A reviewer can only check what they can see: the clause or authority relied on, the playbook position it departs from, and a link to the source.
Flag uncertainty. Output that says when inputs are unusual or the answer is uncertain gets a closer read than output that always sounds sure.
Match review depth to risk. Ten minutes on a first-draft internal memo and a line-by-line check of every citation in a court filing are both proportionate.
Keep an audit trail. Record what the AI proposed, what the lawyer changed, who approved it and when. That record is your evidence of supervision under Rules 5.1 and 5.3 and paragraph 3.5 of the SRA Code.
Limit what agents can do on their own. If a tool can send email, file documents or update matter records, decide in advance which actions need a lawyer's approval. Anything binding on the client should.

What do MENA regulators say about human oversight?

There is no Gulf statute that copies Article 14. Saudi Arabia's SDAIA set out human oversight and accountability in its AI Ethics Principles (September 2023), which are guidance. In the DIFC, Regulation 10 of the Data Protection Regulations covers personal data processed through autonomous and semi-autonomous systems, including AI. Bar and professional rules across the region already hold the licensed lawyer responsible for advice and filings, whatever tool produced the draft. General legal information, not legal advice; check the current rules of your own regulator.

How does HAQQ support human review?

HAQQ is built for lawyers who review before they send. Justinian, HAQQ's legal AI engine, shows the sources behind each answer so they can be checked, and a lawyer signs off. In Legal AI Chat, personal details are replaced with stand-ins before a request reaches an external model and restored before you see the answer. Firms can choose EU, US or Middle East hosting, a private cloud or on-premise deployment, with SSO, role-based access and audit logs.

To start, write down your firm's review rule for each output type (internal note, client email, contract markup, court filing), name who approves each, and test one tool against those rules on a closed matter before it touches a live one. You can run that test in HAQQ Legal AI Chat.

Sources

Key facts

ABA Formal Opinion 512 on generative AI tools, issued 29 July 2024, applies Model Rules 1.1 (competence), 1.6 (confidentiality), 5.1 and 5.3 (supervision) and others, and requires an appropriate degree of independent review of AI output (ABA, 29 Jul 2024).
Article 14 of the EU AI Act (Regulation (EU) 2024/1689) requires high-risk AI systems to be designed so that people can understand, monitor, correctly interpret, override and stop them, and to stay aware of automation bias (EUR-Lex text, read 20 Oct 2026).
Under Regulation (EU) 2026/1744, the high-risk obligations for Annex III systems, Article 14 included, now apply from 2 December 2027 (Official Journal, 24 Jul 2026).
The SRA Code of Conduct for Solicitors keeps solicitors accountable for work done through others (paragraph 3.5) and bound by confidentiality (paragraph 6.3); in R (Ayinde) v Haringey [2025] EWHC 1383 (Admin), the Divisional Court said firm leaders must take practical steps so everyone checks AI research against authoritative sources (Judiciary of England and Wales, 6 Jun 2025).
Damien Charlotin's AI Hallucination Cases database lists 2,202 court decisions involving AI-fabricated or misrepresented material (database, last updated 19 Oct 2026).

Human Oversight Spectrum in Legal AI

Full Automation

0% human

Critical Risk

Human-on-the-Loop

30% human

High Risk

Human-in-the-Loop

70% human

Managed Risk

Human-Led

100% human

Minimal Risk

AI decides and acts without review. No audit trail, no accountability chain. Unacceptable for legal work.

What does human in the loop mean in legal AI?

Human-in-the-Loop Workflow

AI Analyzes

AI Proposes

Lawyer Reviews

Modify or Approve

Final Decision

AI AnalyzesAI processes document and generates output

What does ABA Formal Opinion 512 require?

The ABA Standing Committee on Ethics and Professional Responsibility issued Formal Opinion 512 on 29 July 2024. It applies existing Model Rules to generative AI:

Rule 1.1 (competence): lawyers need a reasonable understanding of what the tool can and cannot do, and must review its output to a degree that fits the task. Relying on output without an appropriate degree of independent verification can breach the duty.
Rule 1.6 (confidentiality): before putting client information into a self-learning tool, which may reuse inputs, the lawyer needs the client's informed consent, and boilerplate in an engagement letter is not enough.
Rules 5.1 and 5.3 (supervision): partners and managers must set clear policies on permitted AI use, train lawyers and staff, and supervise the work, including vendors and their tools.
Rule 1.5 (fees): lawyers can bill for the time they spend, including reviewing AI output, but not for time saved, and not for learning a general-purpose tool unless the client asked for that tool.

Several state bars, including California, Florida and New York, have issued their own guidance with similar review duties. Our bar-rules tracker lists them by country and state.

What does Article 14 of the EU AI Act say about human oversight?

understand the system's capacities and limitations and monitor its operation;
remain aware of automation bias, the tendency to over-rely on output;
interpret the output correctly;
decide not to use the system, or disregard, override or reverse its output;
interrupt it through a stop procedure that leaves it in a safe state.

Even where Article 14 does not bind a firm directly, its list is a useful checklist for designing review: can the reviewing lawyer understand, interpret, override and stop the tool?

What do the SRA and the courts in England and Wales expect?

AI Failure Modes That Only Humans Catch

Hallucinated Citation

AI cites "Section 14.2": which does not exist

Jurisdictional Error

Applies NY law standards to a UAE-governed contract

Context Collapse

Ignores a carve-out that changes the meaning of an obligation

Privilege Breach

Sends privileged strategy notes to a third-party API

False Confidence

Marks a high-risk clause as "standard" with 95% confidence

Human review catches errors that no accuracy metric can prevent

Vyzkoušejte HAQQ AI zdarma

Zažijte sepisování a rešerše s podporou AI

Which AI errors does human review catch?

These are the errors we see most often when lawyers test AI on real matters, and each needs someone who knows the law and the file:

Fabricated or misdescribed authority. A case, section or quotation that does not exist, or a real case cited for a point it does not support. Charlotin's database counts 2,202 court decisions on this. Our hallucination audit breaks the cases down by court and sanction.
Wrong jurisdiction. Common-law reasoning applied to a civil-law contract, or GDPR concepts imported into a Saudi data processing agreement governed by the Saudi Personal Data Protection Law.
Missed interactions. A limitation of liability clause that reads as standard until a carve-out three sections later changes it, or a side letter the AI never saw.
Confidentiality exposure. Client documents sent to a consumer tool whose terms allow it to keep or reuse inputs. In United States v. Heppner (S.D.N.Y., February 2026), documents a defendant generated with consumer Claude were held protected by neither privilege nor work product.
Overconfident output. A clause marked "standard" with no reasoning attached, which invites the reviewer to skip it. Article 14 names this automation bias.

How do you design human review that works?

Put the gate in the workflow. Drafts, risk flags and redlines should need an approval step before they can be sent, filed or applied. An approval button in the tool is harder to skip than a policy people must remember.
Show the reasoning and the sources. A reviewer can only check what they can see: the clause or authority relied on, the playbook position it departs from, and a link to the source.
Flag uncertainty. Output that says when inputs are unusual or the answer is uncertain gets a closer read than output that always sounds sure.
Match review depth to risk. Ten minutes on a first-draft internal memo and a line-by-line check of every citation in a court filing are both proportionate.
Keep an audit trail. Record what the AI proposed, what the lawyer changed, who approved it and when. That record is your evidence of supervision under Rules 5.1 and 5.3 and paragraph 3.5 of the SRA Code.
Limit what agents can do on their own. If a tool can send email, file documents or update matter records, decide in advance which actions need a lawyer's approval. Anything binding on the client should.

Legal AI s člověkem v procesu: co vyžadují ABA 512, AI Act EU a SRA

Key facts

What does human in the loop mean in legal AI?

What does ABA Formal Opinion 512 require?

What does Article 14 of the EU AI Act say about human oversight?

What do the SRA and the courts in England and Wales expect?

Which AI errors does human review catch?

How do you design human review that works?

What do MENA regulators say about human oversight?

How does HAQQ support human review?

Sources

Související zdroje

Související články

Často kladené otázky

Co dál?

Vyzkoušejte HAQQ AI zdarma

Spočítejte si svou ROI

Procházejte více než 600 právních promptů

Legal AI s člověkem v procesu: co vyžadují ABA 512, AI Act EU a SRA

Key facts

What does human in the loop mean in legal AI?

What does ABA Formal Opinion 512 require?

What does Article 14 of the EU AI Act say about human oversight?

What do the SRA and the courts in England and Wales expect?

Which AI errors does human review catch?

How do you design human review that works?

What do MENA regulators say about human oversight?

How does HAQQ support human review?

Sources

Související zdroje

Související články

Často kladené otázky

Co dál?

Vyzkoušejte HAQQ AI zdarma

Spočítejte si svou ROI

Procházejte více než 600 právních promptů