The shelf behind the course.

The scholarship

• May Contain Lies (opens in a new tab)2024

  Alex Edmans

  Source of the Ladder of Misinference (statement → fact → data → evidence → proof): the course-wide decoding tool, installed in Session 2 and drilled in its legal reading in Session 6. A hallucinated citation is a statement costumed as binding proof.

• Thinking, Fast and Slow (opens in a new tab)2011

  Daniel Kahneman

  System 1 and System 2; anchoring (including the sentencing-dice study on judges): the pedigree of Session 2's flagship experiment, run on the room and then on the model.

• “Playing Dice with Criminal Sentences” (Personality and Social Psychology Bulletin) (opens in a new tab)2006

  Englich, Mussweiler & Strack

  The sentencing-dice study described in Thinking, Fast and Slow: judges who rolled a higher number handed down longer sentences. Session 2's anchoring experiment, in the literature.

• “Why Language Models Hallucinate” (opens in a new tab)2025

  Kalai, Nachum, Vempala & Zhang

  The mechanism behind the confidence: models are optimised like test-takers, and under binary grading a confident guess beats abstention, so they answer. Session 2's core citation, and the reason 'permission to say I am not certain' works.

• Prompting Science Reports 1–3 (Wharton) (opens in a new tab)2025

  Meincke, Mollick, Mollick & Shapiro

  The folklore autopsy's evidence: politeness effects unpredictable; chain-of-thought prompting marginal on reasoning models; tipping and threatening do nothing measurable. Session 1.

• “Lost in the Middle: How Language Models Use Long Contexts” (TACL) (opens in a new tab)2024

  Liu et al.

  The position effect behind Session 3's ordering rules: material buried mid-context is found less reliably. Newer models flatten the effect without abolishing it.

• “LLMs Get Lost in Multi-Turn Conversation” (opens in a new tab)2025

  Laban et al.

  An average 39% drop from a clean single ask to a meandering conversation, across 15 models. The evidence behind Session 3's fresh-window habit.

• “Large Language Models Cannot Self-Correct Reasoning Yet” (ICLR 2024) (opens in a new tab)2024

  Huang et al.

  Models struggle to correct their own reasoning without an external signal, and can get worse after self-review. This is why Session 5's critic runs in a fresh context.

• Computational Thinking, CACM 49(3) (opens in a new tab)2006

  Jeannette Wing

  The course's lens: decomposition is issue-spotting, abstraction is extracting what matters, and legal training is therefore a head start at this craft. Sessions 1 and 3.

• Calling Bullshit (opens in a new tab)2020

  Bergstrom & West

  Verification heuristics: who's telling me this, how do they know, what are they selling. Session 6's trace-to-origin reflex, alongside Caulfield's SIFT.

• Noise (opens in a new tab)2021

  Kahneman, Sibony & Sunstein

  Noise in expert judgment (identical cases, different sentences), and the frame for treating one model answer as one sample from a distribution. Background reading.

• Hello World (opens in a new tab)2018

  Hannah Fry

  Algorithms and where they go wrong; the centaur idea: human judgment paired with machine speed. Background reading.

• Weapons of Math Destruction (opens in a new tab)2016

  Cathy O'Neil

  Opaque, biased algorithmic decision systems: the policy backdrop to the draft court Regulations' ban on AI risk-scoring. Background reading.

• You Look Like a Thing and I Love You (opens in a new tab)2019

  Janelle Shane

  The accessible, funny on-ramp to how AI fails. Background reading.

• The Art of Statistics (opens in a new tab)2019

  David Spiegelhalter

  Base rates and conditional probability behind expert evidence. Background reading.

• How to Lie with Statistics (opens in a new tab)1954

  Darrell Huff

  Statistical-misuse source; teach the credibility irony (Huff's tobacco work). Background reading.

• Science Fictions (opens in a new tab)2020

  Stuart Ritchie

  Research-integrity failures; supports the credibility-irony teaching. Background reading.

• “Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools”, Journal of Empirical Legal Studies 22(2) 216 (2025) (opens in a new tab)

  Magesh, Surani, Dahl, Suzgun, Manning & Ho

  The Stanford evaluation: professional grounded legal-research tools still hallucinated on roughly one query in six (Lexis+ AI ≈17%) to one in three (Westlaw AI-AR ≈33%), against ≈43% for 2024-era GPT-4, and against vendor claims of “hallucination-free”. Sessions 4 and 6; teach the vendor contestation too.

• “AI-Powered Lawyering: AI Reasoning Models, Retrieval Augmented Generation, and the Future of Legal Practice” (J. Law & Empirical Analysis) (opens in a new tab)

  Schwarcz, Manning, Prescott, Barry, Cleveland & Rich

  The positive case: in the first randomised controlled trial on upper-level law students, both a grounded legal tool and a reasoning model significantly improved work quality, with productivity up 50–130% on most tasks. Session 4 background.

• “How Much Do Legal RAG Systems Still Hallucinate?” (opens in a new tab)

  Das, Abualhaija & Bianculli

  The 2026 follow-up: across eight legal RAG systems, hallucinations remain pervasive: under 10% for the best, near half for the worst. Grounding raises the floor; it does not end the duty. Session 4 background.

• “Deep Fakes: A Looming Challenge for Privacy, Democracy, and National Security”, 107 California Law Review 1753 (2019) (opens in a new tab)

  Chesney & Citron

  The legal analysis of synthetic media and the source of the “liar's dividend”. Session 7.

• Tomorrow's Lawyers (opens in a new tab)

  Richard Susskind

  The changing legal job: routine work automated, judgment and verification the premium skills. Session 8 background.

• Artificial Intelligence and Legal Analytics (2017) (opens in a new tab)

  Kevin D. Ashley

  What machines can and cannot do with legal texts. Background reading.

• The Black Box Society (2015); New Laws of Robotics (2020) (opens in a new tab)

  Frank Pasquale

  Opacity of algorithmic systems; regulating AI with the human in the loop. Background reading.

• Is Law Computable? (2020) (opens in a new tab)

  Deakin & Markou (eds.)

  On the computability of law; background law-and-technology reading.

Primary cases

1. 01
   Pooja Ramesh Singh v. Jammu and Kashmir Bank Ltd. (opens in a new tab)2026 INSC 668 (SC of India, 2 July 2026, Narasimha & Aradhe JJ.; reportable)

   The course's India-first anchor: citing unverified AI-generated judgments is misconduct for an advocate and a serious lapse for a judge; a decision resting on even “an iota” of hallucinated material is “no decision in the eyes of the law”; NCLT and NCLAT orders set aside; BCI directed to frame guiding principles and disciplinary consequences (Sessions 2, 6 and 8).

2. 02
   Gummadi Usha Rani v. Sure Mallikarjuna Rao (opens in a new tab)SLP (C) No. 7575 of 2026 (SC of India, order of 27 Feb 2026; pending)

   The origin story: on a trial-court order resting on four non-existent SCC citations, the Court observed that such a decision “would be a misconduct and legal consequence shall follow”, issued notice and directed a BCI expert committee. An interim order in a pending matter; the holding now lives in Pooja Ramesh Singh (Session 2).

3. 03
   Pune Bar Association v. Union of India (opens in a new tab)W.P. (C) No. 599 of 2026 (SC of India, 22 May 2026)

   Three-judge order rejecting a constitutional challenge to BSA s.63(4): the certificate's hash-value requirement has a rational nexus with integrity (“an electronic fingerprint”), and Part B's “expert” is not confined to s.79A-notified examiners, expressly noting that authenticity challenges are “accentuated with the advent of artificial intelligence and deepfake technology” (Sessions 4 and 7).

4. 04
   In Re: Summoning Advocates who give legal opinion or represent parties during investigation of cases (opens in a new tab)Suo Motu W.P. (Crl.) No. 2 of 2025, 2025 INSC 1275 (SC, 31 Oct 2025); (2026) 2 SCC 233

   Privilege under BSA ss.132–134 construed: the privilege is the client's, invocable by the advocate; summoning defence counsel sharply confined; device productions court-supervised; in-house counsel outside s.132, with only s.134's narrower protection; the privilege frame under the course's never-paste rule (Sessions 4 and 7).

5. 05

   Anvar P.V. v. P.K. Basheer

   (2014) 10 SCC 473 (SC, three-judge bench)

   The electronic-evidence certificate held mandatory for secondary electronic evidence, overruling Navjot Sandhu. The start of the doctrine line Session 4's workspace is built on, carried into the BSA era as the working assumption (Sessions 4 and 7).

6. 06

   Arjun Panditrao Khotkar v. Kailash Kushanrao Gorantyal

   (2020) 7 SCC 1 (SC, three-judge bench)

   The certificate mandatory for secondary evidence of an electronic record, not required where the original is produced; Shafhi Mohammad (2018) 2 SCC 801 overruled. Session 6's living example of the failure class everyone misses: law that was good until it was not (Sessions 4, 6 and 7).

7. 07
   Deepak v. Heart & Soul Entertainment Ltd. (opens in a new tab)Bombay HC, 7 Jan 2026 (Sathaye J.; 2026:BHC-AS:828)

   ₹50,000 in costs for unverified AI-generated written submissions citing an untraceable judgment. The sanctioned party was a company appearing through its director in person; the warning extends to advocates as dictum (Sessions 2 and 8).

8. 08
   Buckeye Trust v. PCIT (opens in a new tab)ITA No. 1051/Bang/2024 (ITAT Bengaluru; order of 30 Dec 2024, recalled 7 Jan 2025)

   The original order rests on authorities that could not be traced; recalled suo motu within days for “inadvertent errors”. The recall order as reported names neither AI nor any tool; the attribution is the press's inference, and the course says so (Sessions 2 and 8).

9. 09
   Cork & Anor v. Smith (opens in a new tab)[2026] EWHC 1199 (Ch) (22 May 2026)

   The supervision anatomy, cited by the Indian Supreme Court in Pooja Ramesh Singh: a firm's internal AI tool fabricated a rule quotation, its own verification warnings were ignored, supervisors approved unchecked, the court was misled twice; regulator referral. Comparative (England & Wales); Session 8.

10. 10
    Jaswinder Singh v. State of Punjab (opens in a new tab)CRM-M-22496-2022, 2023:PHHC:044541 (P&H HC, 27 Mar 2023, Chitkara J.)

    The court recorded a ChatGPT query on bail jurisprudence under a “post-reasoning” heading, expressly not as a basis for the decision; the Indian example of a court using AI to navigate without treating it as authority (Session 5 background).

11. 11
    Mata v. Avianca, Inc. (opens in a new tab)678 F. Supp. 3d 443 (S.D.N.Y. 2023)

    Six fabricated cases, a $5,000 sanction, and the fatal assumption that a fluent tool “could not possibly be fabricating cases”. The comparative flagship, now one of roughly 1,700 documented decisions worldwide (Sessions 2 and 8).

12. 12
    ANI Media v. OpenAI (opens in a new tab)Delhi HC, CS(COMM) 1028/2024 (Bansal J.; interim injunction refused 24 Jul 2026)

    India's first generative-AI copyright suit over training data. On 24 July 2026 the interim injunction was refused: LLM training was held, prima facie, fair dealing under s.52(1)(a), with the suit continuing to trial and the appeal window open. Background reading; re-verify status before citing.

13. 13
    State v. Loomis (opens in a new tab)881 N.W.2d 749 (Wis. 2016)

    The COMPAS risk-score challenge: the comparative backdrop to the draft Indian court Regulations' absolute ban on AI risk-scoring in bail and sentencing. Comparative (US); Session 7 background.

14. 14
    Da Silva Moore v. Publicis Groupe (opens in a new tab)287 F.R.D. 182 (S.D.N.Y. 2012) (Peck M.J.)

    First judicial approval of predictive coding in e-discovery; AI deployed transparently and defensibly. Comparative (US); background reading.

15. 15
    NYT v. OpenAI / Microsoft (opens in a new tab)S.D.N.Y., No. 1:23-cv-11195 (pending)

    The US counterpart on copyright in training data. Background reading; status may have moved.

Governing instruments

• Bharatiya Sakshya Adhiniyam 2023 — s.63 and the Schedule; ss.132–134

  The electronic-records regime (certificate with Part A hash value and Part B expert signature) and the advocate-client privilege provisions; the statutory spine of Sessions 4 and 7. Verify against the Gazette text before council use.

• IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules 2026 (opens in a new tab)

  G.S.R. 120(E), 10 Feb 2026; in force 20 Feb 2026. India's synthetic-media regime: “synthetically generated information” defined (audio-visual; text excluded), prominent labels and embedded provenance metadata required, takedown clocks compressed to hours. Session 7.

• Supreme Court of India — draft Regulations for Use of Artificial Intelligence in Courts, 2026

  Published for consultation June 2026; a draft as of Aug 2026; re-check notification each cohort. Fifty-seven regulations: permitted judicial uses, an absolute bar on AI deciding or risk-scoring (reg. 20(d)), lawyer disclosure certificates (reg. 43(3)), mandatory AI training for advocates (reg. 49). Session 7.

• Supreme Court of India White Paper on Artificial Intelligence and Judiciary (opens in a new tab)

  Centre for Research and Planning, Nov 2025. A research document, not an operative instrument: recommends secure in-house tools, and states Guidelines 12 (no confidential or privileged information into any AI tool), 14 (independent verification before reliance), 15 (no AI verifying AI) and 18 (responsibility stays human). It contains no cloud restriction and no approved-tools regime; both claims circulate and are wrong. Sessions 5 and 7.

• DPDP Act 2023 + DPDP Rules 2025 (opens in a new tab)

  India's data-protection statute (enacted 11 Aug 2023; Rules notified 13 Nov 2025). The fiduciary/processor frame under the confidentiality discipline; substantive obligations commence 13 May 2027. Sessions 4 and 7.

• EU AI Act (opens in a new tab)

  Regulation (EU) 2024/1689; judicial-use systems high-risk under Annex III point 8, with the application date for those obligations deferred to 2 Dec 2027 by the 2026 omnibus amendment. Comparative contrast in Session 7.

The machine side