LiTL — Lawyer in the Loop

As a small team exploring the intersection of litigation and LegalTech we began with a simple question: when AI helps prepare legal work how does a lawyer actually verify what it produced and how is that verification recorded?

Generative AI has made it remarkably easy to produce a first draft, find authorities and assemble legal material. But the work does not end when the draft is generated. Before something reaches a client, a court or a senior reviewer someone still has to ask a more fundamental question: can we stand behind what this document says?

A draft can cite a case that does not exist. It can cite a real case with the wrong citation. A quotation can read as authoritative without appearing in the judgment. And perhaps most difficult of all the authority may be genuine, correctly named and correctly reported yet still not support the proposition it is cited for. Ordinary citation checking cannot catch that last problem. The question is no longer simply “does this case exist?” but “what does the source actually say and does it support what this draft asks the reader to believe?”

This distinction is becoming increasingly important. In Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. 2026 INSC 668 the Supreme Court dealt with reliance on non-existent fake and hallucinated AI-generated material presented as precedent. While recognising AI as an assistive tool the Court asserted total and absolute human control over adjudication with a human in the loop at every stage. More recently on 2 September 2026 Vijay Ghanshyam Gadiya v. Union of India 2026 INSC 947 brought the issue into even sharper focus. The Court found that some authorities relied upon were non-existent or carried fake citations while others existed but did not lay down the propositions attributed to them. The Court set aside the High Court order and the ₹425.27 crore penalty order and remitted the matter for fresh adjudication.

For us these developments point to a simple distinction at the centre of LiTL: a source being found is not the same as a source being verified.

LiTL — Lawyer in the Loop is built around that gap. It provides a structured verification layer for AI-assisted legal work. It brings the relevant source and the proposition being relied upon into the same workflow and leaves the final assessment where it belongs — with the lawyer.

The lawyer can Confirm, Correct, Reject or Leave Unresolved. Each decision becomes part of a verification record showing what was identified, what was located, what was reviewed, what was changed and what remains open.

We deliberately treat Unresolved as a legitimate outcome. A source may be unavailable, a passage may require deeper reading or an issue may need to be taken to a senior. A system that forces every item into a resolved state may produce a cleaner report but not necessarily a truer one.

LiTL does not certify that a document is legally correct and it does not turn “source found” into “verified.” It preserves the boundary between what the system found, what the source contains and what the lawyer decided.

The Supreme Court’s Draft Regulations for Use of Artificial Intelligence in Courts 2026 envisage that a Court may require disclosure of the AI system used, the extent of its assistance and the steps taken to verify its output. LiTL is designed with this direction in mind. It makes the verification process visible, reviewable and capable of leaving a record.

Not AI versus the lawyer. AI evidence for the lawyer.

Product Note: LiTL is currently presented as a fully interactive verification workflow. Documents can be processed through the verification workflow, relevant sources can be fetched and reviewed alongside the draft, and a verification report is generated from the recorded review. We are continuing to refine the product and expand its broader workflow capabilities.

Detection — Finds citations, statutory references, quotations and other source-dependent claims in an AI-assisted draft.

Location — Locates the underlying authority and places the relevant source passage beside the proposition it is being used to support.

Unsupported-proposition flagging — Surfaces cases and authorities that exist and are correctly cited, but do not sufficiently support the proposition attributed to them. This is the failure ordinary citation checking can miss.

Quotation comparison — Shows the draft wording and source wording side by side, unedited and at the same size.

Repealed-provision mapping — Identifies provisions from repealed statutes and maps them to their successor provisions, such as Section 436A CrPC → Section 479 BNSS, 2023.

Four honest outcomes — Confirm, Correct, Reject, or Leave Unresolved. No score, grade or confidence percentage.

Unresolved as a legitimate state — Open items remain visible and travel with the document to the senior reviewer, rather than being resolved simply for the sake of a clean report.

Verification record — Creates a timestamped record of what was detected, what was located, what the reviewer decided, and what remains open. It is a record of process, not a certificate of correctness.

Senior review — The senior reviewer sees the same verification record as the associate, including open items. Approval records that the review was completed; it does not automatically close an unresolved item.

Model-agnostic — Designed to work with AI-assisted legal output regardless of the underlying AI model or platform.

We would really value feedback from practising lawyers and legal technology teams on the verification workflow and how similar checks are handled in real legal work. We would also be grateful to connect with people who may be interested in helping us test and refine LiTL using real-world legal workflows. As we continue developing the product, guidance on reliable legal-source integration and privacy and confidentiality considerations would also be greatly appreciated.