Scan: Will a court find that OpenAI has infringed the copyright of the New York Times?

← Back to scans

Will a court find that OpenAI has infringed the copyright of the New York Times? — New York Times wins

NYTOAI-27DEC31 · Science and Technology · 2026-05-16

34%

Agent

70%

Market Price

-36.1%

Edge

40%

Confidence

Volume: 51,839

Spread: 0.2c

Days to resolution: 594

Markets in event: 1

Final Rationale

My independent analysis yielded ~14% based on a strict reading: an affirmative SDNY ruling of infringement before Jan 2028, with the case still in discovery as of May 2026 and ~20 months remaining. I PARTIALLY ACCEPT the Devil's Advocate critiques on resolution-criteria ambiguity and underweighted partial-SJ pathway — a Kalshi 'NYT wins' market plausibly resolves YES on partial summary judgment for any count (e.g., DMCA §1202 or output-reproduction), which doesn't require full fair-use adjudication and could occur on a faster timeline. I REJECT the settlement-as-win framing as a primary driver (most settlements are confidential without liability admissions and typically resolve NO), but ACCEPT that resolution ambiguity should widen my distribution. The market at 70% with a sustained +9% upward trend likely reflects either broader resolution interpretation or non-public information from the case (which has had several adverse procedural rulings for OpenAI); I deviate downward because the strict legal pathway to a finding within 20 months remains genuinely difficult, and two recent district-court fair-use rulings favoring AI defendants are real headwinds the market may be underweighting.

Price vs. prediction

How the market moved after the agent's call. Solid blue line is the actual market price; dashed line is the agent's static forecast at $t{=}0$.

$n = 58$ follow-ups

Re-scan Context

This market has been scanned before. Previous predictions:

Date	Predicted	Market Price	Confidence
2026-05-08	30%	69%	45%
2026-05-01	22%	71%	40%
2026-04-05	14%	63%	62%

1. Planner Opus · Fermi decomposition + tool plan

Tools chosen: kalshi_data court_docket article_search web_search wikipedia_lookup code_execution

Sub-questions (Fermi decomposition)

Will the NYT v. OpenAI case proceed to trial (rather than settle) before January 2028? (w=0.35)
If the case proceeds to trial, will the Southern District of New York find OpenAI liable on at least one copyright count before January 2028? (w=0.35)
Will OpenAI's fair use defense be rejected by the court (either at summary judgment or trial) before January 2028? (w=0.30)

Planner reasoning

This question resolves YES only if the SDNY affirmatively finds OpenAI liable for copyright infringement (or any related count) before January 1, 2028 — approximately 594 days away. The analysis has several distinct layers: **Timeline dynamics**: The NYT filed suit in December 2023. As of May 2026, roughly 2.5 years have elapsed. Complex IP cases in SDNY typically take 3-5 years to trial. Discovery in a case of this complexity (involving massive AI training datasets) is likely still ongoing or recently completed. A trial before January 2028 is possible but requires rapid progression through pretrial motions and scheduling. **Settlement vs. trial**: The dominant outcome in high-profile media-AI copyright cases is settlement. Major publishers (News Corp, Axel Springer) have settled rather than litigated. The NYT has been notably more adversarial, but settlement pressure grows as both sides face discovery costs. ~75-85% probability of settlement before trial, which would NOT resolve this market as YES (settlements don't constitute a court finding). **Fair use defense strength**: OpenAI's primary defense is fair use — that training on text is transformative and non-substitutive. The law here is genuinely unsettled. The 'intermediate copying' doctrine and 'transformative use' precedents (Authors Guild v. Google) provide some support for OpenAI, but the NYT's evidence of near-verbatim memorization undermines the non-substitution prong. Courts could go either way. **Reference classes**: In major copyright litigation (music, film, publishing), plaintiffs win roughly 40-55% of cases that actually reach verdict. However, AI training data cases are novel — no binding precedent exists. The court may also resolve via summary judgment on discrete counts. **What needs to happen for YES**: (1) No settlement, (2) Pretrial motions resolved without dismissal, (3) Trial scheduled and completed before late 2027, (4) Court rules for NYT on at least one count. Each step has meaningful failure probability. The joint probability is quite low (~5-12%) given the tight timeline and settlement likelihood.

2. Research Parallel tool calls · legacy v2

Tool	Status	Time	Items
kalshi_data	OK	1.8s	-
court_docket	OK	1.2s	-
article_search	OK	0.7s	-
web_search	OK	52.0s	-
wikipedia_lookup	OK	1.6s	-
code_execution	OK	0.0s	-

3. Evidence Synthesis Sonnet · 14 items · legacy v2

#	Strength	Credibility	Direction	Source	Claim	Priced In
1	STRONG	72	↓ DOWN	web_search	As of early 2026, no firm trial date has been publicly scheduled for NYT v. OpenAI; the case remains in active discovery/pretrial proceedings with disputes over expert depositions and ChatGPT prompts as recently as March 2026.	Yes
2	MODERATE	68	↓ DOWN	web_search	The NYT case is part of consolidated MDL pretrial proceedings involving 16 copyright lawsuits in SDNY, adding complexity that typically extends litigation timelines significantly.	Yes
3	MODERATE	65	↓ DOWN	article_search	The majority of high-profile media-AI copyright disputes (News Corp, Axel Springer) have settled rather than gone to trial; the NYT has been more adversarial but settlement pressure grows with discovery costs.	Yes
4	WEAK	35	↓ DOWN	code_execution	A base-rate model estimates P(no settle) at ~20% and P(trial completes by Jan 2028 \| no settle) at ~40%, yielding a ~8% overall probability that NYT wins via trial or summary judgment.	No
5	MODERATE	55	↑ UP	kalshi_data	The Kalshi market 'New York Times wins' (NYTOAI-27DEC31) is currently priced at 70.1%, up ~9.1% over 30 days, with moderate volume (~107 contracts/day), suggesting the crowd assigns meaningful probability to a YES outcome.	Yes
6	MODERATE	65	↓ DOWN	web_search	OpenAI cites two separate federal judges who independently found AI model training 'highly transformative and protected by fair use,' suggesting judicial tailwinds for OpenAI's defense.	No
7	MODERATE	62	↑ UP	web_search	Discovery has centered on 'regurgitation' — evidence that ChatGPT memorized and reproduced NYT articles near-verbatim — which undermines the non-substitution prong of OpenAI's fair use defense and strengthens the NYT's case if it reaches trial.	No
8	STRONG	75	NEUTRAL	web_search	No binding precedent exists on AI training data and copyright; the legal landscape is genuinely unsettled, creating high uncertainty in any jury or bench verdict.	Yes
9	WEAK	35	NEUTRAL	code_execution	Base rate for plaintiff winning in major copyright cases that reach trial is estimated at ~45%; combined with the low probability of trial occurring, joint probability of NYT win via trial is ~3.6%.	No
10	STRONG	70	↓ DOWN	web_search	Two federal judges in separate AI copyright cases have independently found that AI model training is 'highly transformative and protected by fair use,' providing precedential support for OpenAI's defense.	No
11	MODERATE	62	↑ UP	web_search	The 'regurgitation' evidence — ChatGPT reproducing verbatim NYT content — directly challenges the 'market harm' and 'substitution' prongs of fair use, potentially making fair use rejection at trial more likely than at summary judgment.	No
12	WEAK	45	↓ DOWN	web_search	OpenAI argues the Times uses 'aggressive and unreasonable litigation tactics' and that the law around AI and fair use has been clarifying in OpenAI's favor, suggesting OpenAI believes the fair use defense is strengthening.	Yes
13	STRONG	72	↓ DOWN	web_search	The case is still in active discovery as of March 2026, meaning summary judgment motions on fair use are likely at minimum 12-18 months away, making pre-January 2028 resolution extremely tight.	Yes
14	MODERATE	50	↑ UP	kalshi_data	The Kalshi market price has risen from a low of 45% to 70.1% over its 151-day history, indicating growing market confidence in a YES outcome, though this conflicts sharply with base-rate analysis.	Yes

Information Gaps

No court docket data was retrieved — actual case scheduling orders, discovery deadlines, or any trial scheduling conferences are unknown.
No information on whether settlement negotiations between NYT and OpenAI are ongoing or have recently broken down.
The identities and reasoning of the two federal judges cited as finding AI training 'transformative' are not specified — whether these are binding or persuasive to SDNY is unclear.
No data on the assigned SDNY judge's views on AI copyright or fair use, or their typical case management timeline.
No information on whether NYT's expert witness dispute (March 2026) was resolved or how it affects the discovery timeline.
No data on whether OpenAI has filed or intends to file a motion to dismiss or early summary judgment motion on any counts.
The Kalshi market resolution criteria are not clearly specified — it's uncertain whether the market correctly reflects the question's narrow YES condition (affirmative court finding, not just settlement).

Key Uncertainties

Whether the NYT and OpenAI will settle before trial — this is the dominant outcome driver; settlement would resolve the Kalshi market as NO.
Whether the SDNY case will be trial-ready before January 2028 given it's still in discovery as of March 2026 with no trial date set.
How the SDNY judge will rule on OpenAI's fair use defense — two other federal judges have favored OpenAI, but the 'regurgitation' evidence may distinguish this case.
Whether the Kalshi market at 70.1% is mispriced due to misunderstanding of resolution criteria (i.e., conflating 'NYT wins' broadly with any legal development vs. an actual court finding of liability).
Whether MDL consolidation will accelerate or delay the NYT's specific case, as consolidated proceedings can cut both ways.
The trajectory of broader AI copyright jurisprudence — new rulings in related cases could create precedent that effectively resolves fair use questions before NYT's case reaches trial.

Market Context

Price: 0.701 · 7d: +0.011 · 30d: +0.091
Volume: MODERATE — 107 contracts/day average, suggesting meaningful but not deep liquidity · Depth: UNKNOWN — no orderbook data provided
Reliability: LOW-MEDIUM — The 70.1% market price appears significantly disconnected from base-rate analysis (~8-10% implied by legal timeline and settlement probability). This discrepancy may indicate: (1) market participants misunderstand resolution criteria, (2) the market is pricing in settlement announcements or partial outcomes incorrectly, or (3) the market has information not reflected in public sources. The 30-day upward trend of +9.1% is notable but the price range history (45%-77%) shows high volatility. This market should be treated with caution as a probability signal.

4. Independent Forecaster Opus · No market price · legacy v2 Did NOT see market price

Independent estimate: 14% (55% confidence)

base rate

16%

55% conf

evidence driven

14%

55% conf

contrarian

12%

55% conf

Spread

4pp

Low disagreement

sq1: Will the NYT v. OpenAI case proceed to trial (or dispositive ruling) before January 2028?

28%

Base rate for complex IP cases reaching trial within ~4.5 years of filing is ~25-35%. The case was filed Dec 2023; by Jan 2028 it will be ~4 years old. As of May 2026, still in active discovery with no trial date set and MDL consolidation adding delay. NYT has been notably adversarial (unlike News Corp/Axel Springer settlements), pushing probability up. However, summary judgment on discrete counts (e.g., verbatim regurgitation) is plausible and could count as a 'court finding' — this broadens the path. Net estimate: ~28%.

sq2: If adjudicated, will SDNY find OpenAI liable on at least one count?

55%

Base rate for plaintiffs winning at least one count in major copyright cases that reach verdict is ~45-55%. The 'regurgitation' evidence (verbatim ChatGPT output of NYT articles) is unusually strong for plaintiffs and is hard to defend under fair use's market-substitution prong. However, two recent federal rulings (Bartz v. Anthropic, Kadrey v. Meta) found training itself transformative, providing OpenAI headwinds on the broader training claim. Winning on 'at least one count' is easier than winning on all — the DMCA or output-reproduction counts are particularly viable for NYT. Estimate: ~55%.

sq3: Will OpenAI's fair use defense be rejected before January 2028?

22%

This requires both adjudication AND rejection. Two recent district court rulings have favored AI training as fair use, creating persuasive (not binding) headwinds. But the regurgitation evidence distinguishes NYT's case. Discovery still active as of March 2026, making SJ motions 12-18 months out — very tight for pre-Jan 2028 resolution. P(SJ or trial reaches fair use ruling) ~30%, P(rejected | reached) ~50% (the output reproduction claim is the most defensible plaintiff path). Estimate: ~22% (with high uncertainty pulling toward base rate).

Show full reasoning trace

base rate: Complex IP cases filed in SDNY reach a dispositive ruling within 4 years roughly 25-35% of the time. Plaintiffs win at least one count in copyright cases that reach verdict ~45-55% of the time. Joint base rate: ~12-18%.

evidence updates: DOWN: No trial date set as of March 2026, still in active discovery, MDL consolidation adds delay (-5%). UP: NYT is unusually adversarial and well-resourced, regurgitation evidence is strong (+3%). DOWN: Two recent federal rulings favoring AI training as fair use create persuasive precedent (-3%). UP: Possibility of partial summary judgment on discrete verbatim-output counts provides an additional path to YES (+2%).

combination method: Conditional: P(adjudication by Jan 2028) × P(NYT wins ≥1 count | adjudication) ≈ 0.28 × 0.55 ≈ 0.15-0.17.

final: My estimate is ~16%. The dominant driver is timeline: the case is still in discovery with no trial date, and ~20 months remain until the deadline. Even with the regurgitation evidence favoring NYT on the merits, the probability of an affirmative court finding (not settlement, not dismissal, not still-pending) before Jan 2028 is low. Confidence is moderate (~0.55) given absence of docket data and settlement intel.

ensemble: {'ensemble_count': 3, 'perspectives': {'base_rate': {'p_yes': 0.16, 'confidence': 0.55}, 'evidence_driven': {'p_yes': 0.14, 'confidence': 0.55}, 'contrarian': {'p_yes': 0.12, 'confidence': 0.55}}, 'spread': 0.04, 'individual_p_yes': {'base_rate': 0.16, 'evidence_driven': 0.14, 'contrarian': 0.12}}

5. Devil's Advocate Sonnet · legacy v2 Market price revealed

Missing Info Reasoning Flaw Asymmetric Evidence

Challenges

Resolution criteria ambiguity: The forecaster assumes a narrow reading ('affirmative court finding before Jan 2028'), but the Kalshi market title is 'New York Times wins.' If the market resolves YES on settlement-with-payment, consent judgment, default judgment, or any partial SJ ruling favoring NYT, the forecaster's narrow interpretation systematically understates p_yes. This information gap is acknowledged but not propagated into uncertainty — the forecast should widen substantially to account for resolution-criteria risk.
Conditional probability composition appears to double-count delay risk. sq1 (28% reaches adjudication) already bakes in timeline risk, and sq2 (55% wins ≥1 count) is conditional on adjudication. The 0.28 × 0.55 = 0.154 math is fine, but sq3 is presented as an independent check rather than a sub-component, creating confusion about whether fair-use rejection is required for YES. If 'wins ≥1 count' on a DMCA or output-reproduction theory doesn't require fair-use rejection, sq3 is largely redundant and shouldn't drag the estimate down.
Partial summary judgment pathway is underweighted. The forecaster notes SJ on discrete counts is 'plausible' but only adds +2%. Given 20 months remain and that SJ motions on narrow output-reproduction/DMCA counts could come well before trial, this pathway alone could plausibly carry 15-25% probability of a NYT-favorable ruling, not 2%.
Market divergence of 56 points with a moderate-volume market trending upward over 30 days deserves more serious engagement. The forecaster dismisses the market as possibly misreading resolution criteria, but provides no evidence for this. If the market is interpreting 'NYT wins' more broadly (including settlements with admission/payment), the forecaster's narrow framing may be the error, not the market's.
Settlement-as-win scenario not modeled. Many high-profile cases (the forecaster notes News Corp, Axel Springer) settle — but a settlement with payment to NYT could plausibly resolve a 'NYT wins' market YES depending on criteria. This isn't addressed.

Suggested adjustment: +8pp

Modest upward adjustment warranted due to (1) underweighted partial-SJ pathway, (2) unmodeled settlement-as-win scenario, and (3) resolution-criteria ambiguity that should widen uncertainty toward the base rate of the market interpretation. Not adjusting fully toward 70% because the forecaster's narrow legal analysis is sound IF the strict interpretation holds, and the recent transformative-use rulings are genuine headwinds. The adjustment reflects reasoning gaps, not market conformity.

Market comparison: Forecaster: 14%, Market: 70%. The 56-point divergence is large. The forecaster's strict reading of 'court finding of infringement before Jan 2028' is legally defensible and the timeline analysis is rigorous. However, the market may be pricing a broader interpretation (settlement, partial ruling, any adverse finding) and the forecaster's failure to resolve this ambiguity is a real gap. The market's steady upward trend (+9% over 30d) on moderate volume suggests genuine information flow, not just noise.

6. Aggregator Opus · final calibrated forecast

34% (40% confidence)