Technology-Assisted Review (TAR)

TAR (predictive coding) uses AI to classify documents in e-discovery. Instead of 20 lawyers reviewing 500,000 documents, the AI learns from a sample and classifies the rest. 70-90% reduction in human review volume.

Process

1. Seed set: senior lawyer manually reviews 200-500 documents (relevant/not relevant/privileged)
2. Training: AI learns classification patterns
3. Bulk classification: AI classifies remaining documents
4. QA: random sample review to validate accuracy
5. Iteration: if accuracy is insufficient, add training data and repeat

TAR 2.0 (Continuous Active Learning)

Evolution: AI prioritizes documents for human review in real time, the most "informative" first. The model improves continuously. More efficient than TAR 1.0 (batch).

Judicial acceptance

• USA: Da Silva Moore v. Publicis (2012), Rio Tinto (2015): TAR can be more accurate than manual review
• UK: Pyrrho Investments (2016)
• EU: no leading case, but procedural rules do not prohibit it. Acceptance trend growing

Document review at scale with generative AI

Beyond TAR, LLMs transform document review:

Modern document review workflow

1. INGESTION: collect docs + OCR + deduplication
2. CLASSIFICATION: TAR relevance + LLM summary/entities/topics
3. HUMAN REVIEW: prioritized by score, enriched with AI summary
4. PRODUCTION: automated privilege log + Bates stamping + export

Document review prompt

Analyze this document and extract:
1. TYPE: contract/email/memo/report
2. DATE and PARTIES mentioned
3. SUMMARY: 2 sentences max
4. RELEVANCE to [case subject]: high/medium/low
5. PRIVILEGE: possibly privileged? Why
6. RED FLAGS: problematic clauses or data
7. AMOUNTS mentioned

If you cannot determine something, mark [VERIFY]. Do not invent.

Case law analysis with AI

Valid use cases

• Semantic search: "rulings on employer liability for AI in hiring" in Westlaw or LexisNexis
• Comparison: upload your brief and ask it to identify case law that supports or contradicts it
• Trends: evolution of interpretation of an article over the last 5 years
• Comparative jurisdiction: how is X resolved in the UK vs Germany

Tools

• Westlaw Edge: semantic search, citation analysis
• LexisNexis: comprehensive case law database, AI-powered research
• Claude/ChatGPT: useful for analyzing a legal text that YOU upload. NOT reliable for citing case law

Predicting litigation outcomes

AI for estimating probability of success based on court history, judge, case type. Useful as additional input for: deciding whether to litigate, estimating provisions, negotiating settlements.

Limitations

• Does not predict legislative or case law changes
• Each case has unique facts the model does not capture
• Historical bias: reproduces past inequalities of the system
• Low accuracy in complex or unprecedented cases

Correct usage: as supplementary information, not as a substitute for professional judgment.

Deepfakes and digital evidence

AI generates audiovisual content indistinguishable from real content. Implications for evidence:

• Questionable authenticity: how to prove a recording is genuine
• Reasonable doubt: the existence of deepfakes allows any digital evidence to be challenged
• Manipulation: parties altering recordings with AI

Mitigation

• Complete digital chain of custody
• Original metadata (EXIF)
• Provenance (C2PA): cryptographic signature at time of capture
• Forensic expert analysis for critical evidence

The Mata v. Avianca case

2023, New York. Lawyer Steven Schwartz filed a brief with 6 case law citations generated by ChatGPT. None existed. Fabricated rulings with perfect formatting: parties, court, date, docket number. All fake.

Consequences

• $5,000 fine for the lawyer
• Worldwide professional embarrassment
• Multiple courts issued Standing Orders requiring disclosure of AI use

Lessons

1. Never cite without verifying. LLMs fabricate with total confidence
2. The duty of diligence is the lawyer's. "The AI generated it" is not an excuse. You sign the brief
3. AI for analysis, not for citation. Excellent at analyzing texts you give it. Dangerous generating references

Ethical obligations of lawyers using AI

• Competence: understanding AI limitations is a professional obligation. Using AI without understanding hallucinations = negligence
• Supervision: AI is an associate that never fact-checks. Output review is mandatory
• Confidentiality: uploading client documents to ChatGPT free tier violates duty of confidentiality. Team/Enterprise plans only
• Transparency: growing trend of courts requiring disclosure of AI use in filings

Workflow for litigation teams

1. RESEARCH: specialized databases (Westlaw, LexisNexis). LLM only for
   analysis of your own texts. NEVER cite without verifying

2. DOCUMENT REVIEW: TAR + LLM summaries. Human oversight on critical docs

3. DRAFTING: AI generates drafts. Senior reviews ALWAYS before signing.
   Verify citations, figures, facts

4. STRATEGY: AI analyzes opposing arguments. Devil's advocate.
   Prediction as input, not as decision

5. COMPLIANCE: verify required disclosure. Enterprise plans with
   data isolation only. Record usage in the case file

Ejercicio practico

Puntos clave