The International Legal Finance Association has filed its response to the UK Government's consultation on competition redress, arguing that reforms intended to speed up the opt-out collective actions regime must not add cost or complexity that makes meritorious claims harder to fund.

The submission responds to the Department for Business and Trade's consultation on "Swifter and Simpler Competition Redress, Regulatory Appeals and Competition Enforcement," which opened on 17 July and closed on 25 September. ILFA's central argument is that the Competition Appeal Tribunal and the appellate courts have already developed workable mechanisms for overseeing class representative suitability, and that the Government should deliver clarity through guidance and the formalisation of existing practice rather than new statutory or procedural requirements.

"Third-party litigation funding is the cornerstone of the opt-out collective actions regime," said Neil Purslow, Chairman of the Executive Committee of ILFA. "Without it, consumers and small businesses would have no realistic means of bringing meritorious claims against well-resourced defendants. In our response, we make it clear that any new reforms must not inadvertently introduce cost or complexity, which only serve to make valid claims harder to bring."

ILFA ties the Government's proposal to permit damages-based agreements in collective proceedings to the unresolved question of funder returns. "Crucially, the Government's proposal to permit damages-based agreements in collective proceedings underscores the urgent need to reverse the PACCAR ruling retrospectively," Purslow said. "To keep this regime viable and investable, we must give funders earlier certainty over returns and introduce better cost budgeting to rein in unpredictable, disproportionate costs."

On costs, the association supports mandatory costs budgeting for claimants and defendants alike from certification onwards, and greater use of alternative dispute resolution where it is required early and backed by real costs sanctions. It also backs a central CAT website for claims and settlements, while cautioning that efficiency measures such as reduced panel composition may yield only marginal savings.

"Maintaining a true equality of arms is essential," Purslow said. "Large defendants should not be allowed to weaponise structural hurdles to quash meritorious claims and ordinary businesses and consumers must remain empowered to hold the powerful to account."

Law firms, funders and legal departments are being sold AI for contract review, legal research and citation checking, and the models change every few months. So we looked only at independent studies published in 2026 that tested the current generation of models from OpenAI, Anthropic and Google on real legal tasks. The short version: the best models are now genuinely good at reading and extracting from documents you give them, still unreliable at recalling law from memory, and the commercial legal research tools lag behind the best custom systems.

At a Glance

Best overall model for legal document work: Google's Gemini 3.1 Pro. It was at or near the top in every 2026 study that tested it, and it was usually the fastest and cheapest of the leaders. OpenAI's GPT-5.5 found slightly more errors in contract review, and Anthropic's Claude models were the most careful about not flagging problems that weren't there.

Best accuracy recorded on a full legal task: 92%, on a 50-state statutory research test run by Stanford, achieved by a purpose-built research tool. The lesson is that how the AI is set up matters as much as which model sits underneath it.

Range of accuracy: from under 7% (asking a model to recall exact case citations from memory) to 99–100% (catching a citation to the wrong case when the model can read the source). Most real-world document tasks landed between 60% and 85%.

Westlaw and Lexis AI: 58% and 64% accuracy on a Stanford statutory survey test, below a custom-built tool at 83–92%.

Biggest single improvement: giving the model the actual documents instead of asking it from memory cut fabricated citations from roughly 15–40% to about 4–15%, and to under 0.2% with a well-built retrieval system.

Key Takeaways

The model to use. For contract review and extraction, start with Gemini 3.1 Pro. It matched the top performer on catching contract errors (74% vs. 75%) at about one-seventh of the cost and in 90 seconds instead of nine minutes. If catching every possible issue matters more than time or cost, GPT-5.5 with reasoning turned on found the most. For checking citations in a brief, the best 2026 results came from GPT-5 running as an agent and from Claude Code with Claude Opus, which was the most precise.

How to Use It

• Give it the documents. Never ask a model to supply case law or citations from memory.
• Turn on the model's "reasoning" or "thinking" mode for review work. It improved error-catching by 9 to 11 points in contract review.
• Use it as a first pass and a second reviewer, not the final reviewer.
• For research, use a tool that pulls from a full, current database of the law, because weak retrieval, not the model, causes many of the errors.

What to Expect

• Contract extraction (pulling out dates, parties, termination and liability terms): about 80–84% accuracy for the best models.
• Final contract proofreading (defined terms, cross-references, inconsistent language): the best models catch about three-quarters of errors. On a 60-page agreement, expect it to miss some.
• Citation checking: nearly all citations to the wrong case get caught, but wrong pinpoint pages slip through 20% to 60% of the time.
• Research answers grounded in the right documents: roughly 6–11% of answers still contain an unsupported statement.

What to Look Out For

• Citations from memory. When asked to recall exact citations without sources, even the best model scored under 7 out of 100, and 20 of 21 models gave confident, wrong answers more than 94% of the time.
• Right case, wrong page. Models tend to approve a citation because the case is on the right topic, even when the cited page doesn't support the point.
• Questions with a false premise. If your question assumes something that isn't true, models often go along with it.
• Legal research tools' marketing. Westlaw AI and Lexis+ AI trailed a custom-built tool by 19 to 25 points on a Stanford test.
• Studies funded by vendors. Some of the best-looking results come from companies selling legal AI. Check who ran the test.

Best Practices

• Ground every task in source documents, and require the model to quote the passage it relied on.
• Check every citation yourself at the pinpoint page before filing. Automated checkers help but don't replace this.
• Turn on reasoning mode for review tasks and accept that it's slower.
• Test a model on a few of your own documents before rolling it out. Rankings change by task.
• Re-test when a new model version arrives; this field moves in months, not years.
• Keep a human reviewer accountable for anything that goes to a court, a client or a counterparty.

What the Studies Found

Contract proofreading. In August 2026, researchers had experienced lawyers plant errors in contracts (misused defined terms, wrong cross-references, wrong party names, contradictions) and tested ten current models on catching them. GPT-5.5 caught 75% of errors, Gemini 3.1 Pro 74%, Claude Sonnet 4.6 69% and Claude Opus 4.7 62%. GPT-5.5 cost $1.38 per contract and took about nine minutes; Gemini 3.1 Pro cost $0.19 and took about 90 seconds. Turning on reasoning mode added 9 to 11 points. Every model was far cheaper than a lawyer, and none was close to perfect.

Contract extraction. A May 2026 study tested models on pulling 26 standard fields out of contracts. Among the major models, Gemini 3.1 Pro scored highest (82%), with Claude Opus 4.6 (82%) and Claude Sonnet 4.6 (80%) close behind and GPT-5.4 at 78%. A smaller legal-specific model built by the study's authors scored 84% at far lower cost. The authors work for Onit, which makes that model.

Made-up citations and facts. A January 2026 study had expert reviewers check 2,700 legal answers from 12 models. Asked without source documents, the best models (GPT-5.2 and Gemini 3.0 Pro) cited something false about 15–17% of the time, and the worst over 30%. Giving the models the relevant documents cut that to about 4–15%. A more carefully built retrieval system brought it below 0.2% for every model.

Research with sources. A March 2026 study found that when models answer from retrieved legal texts, Gemini 3.1 Pro produced unsupported statements 5.7% of the time versus 11.3% for GPT-5.2, and that the quality of the search step mattered more than the choice of model. Its authors sell the search component that performed best. An August 2026 study of eight research setups found unsupported answers ranging from under 10% for the best to nearly half for the worst, with the worst results on questions built on a false assumption.

Westlaw and Lexis. In a February 2026 Stanford study, researchers tested legal AI tools against a Department of Labor survey of state unemployment insurance laws. Westlaw AI scored 58% and Lexis+ AI 64%, while a custom statutory research tool scored 83%, rising to 92% after the researchers found that some of its "errors" were gaps in the government's own survey.

Citation checking. A June 2026 study found more than 1,000 court filings containing fabricated citations, a number growing every year, and tested AI checkers on catching them. GPT-5, working as an agent that looks up cases, caught 83% of planted errors; Claude Code running Claude Opus 4.8 was the most precise and scored best overall. No model reliably caught wrong pinpoint cites, partly because page numbers often sit behind Westlaw and Lexis paywalls. A separate August 2026 study found models catch 93–100% of citations to the wrong case but miss many citations to the wrong page, and even GPT-5.4 with full reasoning missed 40% of wrong pinpoints in court opinions.

Citations from memory. A May 2026 study built from 1,000 real U.S. judicial opinions asked 21 models to recall exact case citations without any sources. The best, Claude Sonnet 4.5, scored under 7 out of 100.

The Bottom Line

The 2026 research is consistent: today's best models, led by Gemini 3.1 Pro, GPT-5.5 and Claude, are useful and cheap for first-pass contract review and extraction when they work from the documents in front of them. They still invent law when asked from memory and still miss wrong pinpoint citations, so a lawyer has to verify anything that leaves the building.

Sources (All 2026)

• Bang et al., "ContractScrub: A benchmark for final review of legal contracts" (Aug. 2026), arXiv:2608.20204
• Lincoln et al., "A Few Good Clauses: Comparing LLMs vs Domain-Trained Small Language Models on Structured Contract Extraction" (May 2026), arXiv:2605.05532
• Dantart, "Reliability by design: quantifying and eliminating fabrication risk in LLMs" (Jan. 2026), arXiv:2601.15476
• Butler and Butler, "Legal RAG Bench: an end-to-end benchmark for legal RAG" (Mar. 2026), arXiv:2603.01710
• Das et al., "How Much Do Legal RAG Systems Still Hallucinate?" (Aug. 2026), arXiv:2608.14210
• Afane et al., "Benchmarking Legal RAG: The Promise and Limits of AI Statutory Surveys" (Feb. 2026), arXiv:2603.03300
• Liu, Stammbach and Henderson, "Who Checks the Citations? Benchmarking Legal Hallucination Detection" (June 2026), arXiv:2606.21155
• Verma, "Is this Citation on Point?" (Aug. 2026), arXiv:2608.12571
• Chen et al., "LegalCiteBench: Evaluating Citation Reliability in Legal Language Models" (May 2026), arXiv:2605.10186