← 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 · 2026-05-08
30%
Agent
69%
Market Price
-39.1%
Edge
45%
Confidence
Volume: 51,389
Spread: 0.1c
Days to resolution: 602
Markets in event: 1
Final Rationale
I accept the Devil's Advocate's critiques on settlement (adversarial discovery posture genuinely reduces settlement probability below the 60-65% base rate for commercial copyright cases), fair use weakness for output reproduction (verbatim ChatGPT outputs of NYT articles are materially harder to defend than training-data fair use, and the N.D. Cal. Anthropic ruling addressed training, not output), and the partial-SJ-ruling asymmetry (a partial liability finding for NYT likely resolves YES, so rounding down was wrong). Adjusting sq1 to ~0.45 and sq2 to ~0.62 yields ~0.28, which I round to 0.30. I reject deferring to the 0.69 market price: liquidity is thin (107 contracts/day, shallow book), the 30-day run-up is noisy, and the gap between this market and base rates strongly suggests it's pricing 'NYT wins eventually' rather than the specific Jan 2028 deadline — a known framing error in thin prediction markets on legal questions with long tails. My 0.30 reflects the genuine strength of NYT's verbatim-reproduction evidence and active SJ schedule while preserving timeline risk that the market appears to ignore.
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:
DatePredictedMarket PriceConfidence
2026-05-01 22% 71% 40%
2026-04-05 14% 63% 62%
1. Planner Opus · Fermi decomposition + tool plan
Tools chosen: article_search web_search court_docket kalshi_data article_search wikipedia_lookup code_execution
Sub-questions (Fermi decomposition)
  1. Will the NYT v. OpenAI case proceed to a merits ruling (rather than settling) by January 2028? (w=0.35)
  2. If the case proceeds to a merits ruling, will the court find OpenAI liable on at least one copyright count before January 2028? (w=0.40)
  3. Will OpenAI's fair use defense fail to shield it from liability in the NYT case? (w=0.25)
Planner reasoning
This question resolves YES only if the Southern District of New York affirmatively finds OpenAI liable on at least one copyright count by January 1, 2028, roughly 602 days away. The key drivers are: (1) Settlement probability: high-profile tech/media copyright disputes settle frequently (~65%), especially when both parties have commercial incentives. Several AI companies have already reached licensing deals with publishers. (2) Timeline to merits ruling: NYT filed in December 2023. By Jan 2028 the case will be ~4 years old. SDNY complex tech copyright cases rarely reach trial judgment in under 4 years, but summary judgment is possible. The case must survive motions to dismiss, complete fact discovery, and briefing within this window. (3) Merits of fair use: The central legal question is whether training LLMs on copyrighted text constitutes fair use. This is genuinely novel law with no circuit precedent. NYT's verbatim reproduction evidence strengthens its case beyond typical training-data arguments. (4) The resolution criterion only requires liability on ANY count, which lowers the bar somewhat. The dominant path to YES is a pre-trial summary judgment ruling rather than a full trial verdict. Reference class: major media vs. tech copyright suits (Authors Guild v. Google took ~10 years; Viacom v. YouTube ~6 years) suggest this timeline is very tight.
2. Research Parallel tool calls · legacy v2
ToolStatusTimeItemsSummary
article_search OK 1.1s -
web_search OK 42.5s -
court_docket OK 0.5s -
kalshi_data OK 1.9s -
wikipedia_lookup OK 1.8s -
code_execution OK 0.0s -
3. Evidence Synthesis Sonnet · 17 items · legacy v2
#StrengthCredibilityDirectionSourceClaimPriced In
1 STRONG 88 ↑ UP web_search In March 2025, Judge Sidney Stein denied OpenAI's motion to dismiss, allowing the main copyright infringement claims to proceed to trial, though no trial date has been set. Yes
2 STRONG 85 ↑ UP web_search Summary judgment is scheduled for April 2026, with discovery still ongoing as of early 2026 — indicating the case is on an active litigation track rather than settlement track. No
3 STRONG 87 ↑ UP web_search On January 5, 2026, Judge Stein affirmed an order compelling OpenAI to produce 20 million anonymized ChatGPT conversation logs — a significant discovery win for the NYT suggesting the case is proceeding aggressively toward merits. No
4 STRONG 82 ↓ DOWN article_search Anthropic settled a major AI copyright lawsuit with authors for $1.5 billion in September 2025, demonstrating that high-profile AI copyright cases can settle before merits rulings even when legally significant. Yes
5 MODERATE 75 ↑ UP web_search OpenAI is contesting multiple discovery orders and fighting preservation requirements, including a planned appeal of the data preservation order, suggesting adversarial dynamics that reduce near-term settlement likelihood. Yes
6 MODERATE 45 ↓ DOWN code_execution A probabilistic model estimates only ~19% chance the case reaches a merits ruling by January 2028, factoring in settlement probability (~65%), timeline pressure, and case complexity. No
7 MODERATE 70 ↓ DOWN web_search Reference class data: Authors Guild v. Google took ~10 years and Viacom v. YouTube ~6 years to resolve, suggesting the Jan 2028 deadline (~4 years from filing) is a very tight window for a merits ruling. Yes
8 STRONG 85 ↓ DOWN article_search In June 2025, a federal judge in San Francisco ruled in favor of an AI company in a first-of-its-kind copyright decision, finding that training LLMs on legally obtained copyrighted works can qualify as fair use. Yes
9 MODERATE 72 NEUTRAL article_search Anthropic's settlement in September 2025 came in a case that involved the 'first substantive decision' on fair use in AI, suggesting the legal landscape still highly uncertain with courts reaching varied outcomes. Yes
10 MODERATE 68 ↑ UP web_search The NYT's verbatim reproduction evidence (ChatGPT reproducing full NYT articles) is substantially stronger than typical training-data fair use arguments, potentially distinguishing it from the favorable AI fair use ruling in San Francisco. Yes
11 MODERATE 80 ↑ UP web_search Judge Stein narrowed but did not dismiss the case in March 2025, preserving the main copyright infringement counts, suggesting the court views the claims as legally viable. Yes
12 MODERATE 75 ↑ UP web_search The NYT's January 2026 discovery win compelling 20 million ChatGPT conversation logs suggests the Times may gather strong evidence of verbatim reproduction at scale, strengthening its merits position. No
13 STRONG 82 ↓ DOWN article_search The June 2025 San Francisco fair use ruling established that AI companies can train on legally obtained copyrighted works without infringement — a precedent that could support OpenAI's fair use defense in the NYT case. Yes
14 STRONG 78 ↑ UP web_search NYT's case includes evidence that ChatGPT reproduces verbatim article content, which goes beyond the training-data fair use question and into direct reproduction — a harder fair use argument for OpenAI to win. Yes
15 MODERATE 72 NEUTRAL web_search There is no circuit court or Supreme Court precedent on AI training data and fair use in the Second Circuit (SDNY), leaving the legal question genuinely open for Judge Stein to decide. Yes
16 WEAK 55 NEUTRAL article_search The Supreme Court's March 2026 ruling in Cox Communications v. Sony Music found that ISPs cannot be held secondarily liable for user infringement, but this case is about direct copying — not directly applicable to the NYT v. OpenAI fair use question. Yes
17 MODERATE 55 ↑ UP kalshi_data Kalshi market prices the NYT wins outcome at 69.1%, up +13.1% over the past 30 days, suggesting market participants significantly increased their estimate of NYT winning recently. Yes
Information Gaps
  • No information on whether settlement negotiations are actively occurring between NYT and OpenAI/Microsoft as of May 2026
  • No detailed information on what the April 2026 summary judgment briefing schedule reveals about the case pace — whether it has been filed, extended, or stayed
  • No information on the specific content of the 20 million ChatGPT logs ordered produced and whether they confirm verbatim reproduction at scale
  • No circuit court or Second Circuit precedent directly applicable to AI training fair use, leaving the legal outcome highly uncertain
  • No information on whether OpenAI has pursued or completed any licensing deal with the NYT (as some other publishers have reached)
  • No information on the trial date — whether one has been set following summary judgment schedule
  • Unclear how the San Francisco fair use ruling (N.D. Cal.) would influence or be distinguished by Judge Stein in SDNY
  • No information on whether Anthropic-style settlement pressure (following the $1.5B authors settlement) has influenced NYT's negotiating posture
Key Uncertainties
  • Whether NYT and OpenAI settle before a merits ruling — high-profile tech/media disputes settle frequently, and the Anthropic precedent may push both sides toward negotiation
  • Whether summary judgment briefing in April 2026 leads to a ruling before January 2028, or whether the case extends to trial
  • How Judge Stein will evaluate the verbatim reproduction evidence vs. the broader fair use defense for AI training — genuinely novel legal question
  • Whether the San Francisco AI copyright ruling will influence the SDNY court or be distinguished on the basis of the verbatim reproduction evidence
  • Whether OpenAI's ongoing discovery battles signal a strategy to delay proceedings past the January 2028 resolution deadline
  • The reliability of the Kalshi 69.1% price — the market may be conflating 'NYT wins eventually' with 'NYT wins by January 2028', or may reflect broader settlement expectations
Market Context
Price: 0.691 · 7d: +0.001 · 30d: +0.131
Volume: LOW · Depth: SHALLOW
Reliability: LOW-MEDIUM — low average volume (107 contracts/day), shallow orderbook; significant 30-day price run-up (+13.1%) may reflect new case developments (January 2026 discovery ruling) but thin trading makes this noisy. The 69.1% price appears significantly higher than the base-rate probability model estimate (~9%), suggesting the market may be pricing 'NYT wins eventually' rather than 'NYT wins by Jan 2028', or there is significant mispricing.
4. Independent Forecaster Opus · No market price · legacy v2 Did NOT see market price
Independent estimate: 21% (55% confidence)
base rate
20%
55% conf
evidence driven
21%
55% conf
contrarian
21%
55% conf
Spread
1pp
Low disagreement
sq1: Will the NYT v. OpenAI case proceed to a merits ruling (rather than settling) by January 2028?
38%
Base rate: Major tech/media copyright suits reaching merits ruling within ~4 years of filing is uncommon (~25-30%). Authors Guild v. Google and Viacom v. YouTube took 6-10 years. Settlement rate for high-profile commercial copyright disputes is ~60-65%. Evidence UP: April 2026 summary judgment scheduled (+10-15%) — this is the dominant path to a pre-2028 ruling. The Jan 2026 discovery win and OpenAI's adversarial discovery posture suggest no imminent settlement (+5%). Motion to dismiss survived in March 2025, case is on active track. Evidence DOWN: Anthropic's $1.5B settlement creates pressure for similar resolution (-5%). Discovery still ongoing in early 2026 means SJ may slip. Even with April 2026 SJ briefing, a ruling could take 6-12 months, leaving narrow window. Settlement remains plausible after SJ ruling preview. Net: ~38%.
sq2: If the case proceeds to a merits ruling, will the court find OpenAI liable on at least one copyright count before January 2028?
55%
Conditional on a merits ruling, the question is whether NYT wins ANY count. Base rate for plaintiffs winning at least one count in surviving copyright cases that reach merits is ~50-55%. Evidence UP: Verbatim reproduction evidence is the strongest fact pattern in any AI copyright case — ChatGPT outputting full articles is direct copying, not just training (+10%). Judge Stein already preserved core counts. The 20M ChatGPT log discovery may show systematic reproduction. NYT only needs to win ONE count (low bar). Evidence DOWN: June 2025 N.D. Cal. ruling that AI training is fair use creates persuasive (non-binding) precedent OpenAI will leverage (-8%). Fair use is genuinely novel in 2nd Circuit. Even verbatim outputs have transformative-use defenses. Net: ~55%.
sq3: Will OpenAI's fair use defense fail to shield it from liability?
50%
This overlaps heavily with sq2. Fair use is the central liability defense. Verbatim reproduction is a much harder fair use case than pure training-data arguments — the four factors (especially market harm and amount used) cut against OpenAI when full articles are reproduced. The N.D. Cal. ruling addressed training, not output reproduction, and may be distinguishable. However, courts have been receptive to transformative-use arguments for AI. Genuinely novel legal terrain in 2nd Circuit. ~50%.
Show full reasoning trace
base rate: Major media vs. tech copyright cases reach merits ruling within 4 years ~25-30% of the time. Plaintiff wins at least one count conditional on reaching merits ~50-55%. Combined unconditional base rate: ~15%.
evidence updates: Active April 2026 SJ schedule and recent discovery wins push sq1 up from base ~25% to ~38%. Verbatim reproduction evidence pushes sq2 up modestly, offset by N.D. Cal. fair use precedent. Anthropic settlement creates settlement pressure but adversarial discovery posture suggests no imminent deal.
combination method: Conditional multiplication of sq1 and sq2 (sq3 subsumed in sq2). 0.38 × 0.55 ≈ 0.21, rounded to 0.20 accounting for partial-ruling scenarios that don't establish liability.
final: Approximately 20% probability of an SDNY liability finding by Jan 2028. The dominant blocker is timeline — even with active litigation, getting from SJ briefing in April 2026 to a liability ruling within 20 months while overcoming OpenAI's delay tactics and settlement pressure is hard. Conditional merits prospects are reasonable given verbatim reproduction evidence.
ensemble: {'ensemble_count': 3, 'perspectives': {'base_rate': {'p_yes': 0.2, 'confidence': 0.55}, 'evidence_driven': {'p_yes': 0.21, 'confidence': 0.55}, 'contrarian': {'p_yes': 0.21, 'confidence': 0.55}}, 'spread': 0.01, 'individual_p_yes': {'base_rate': 0.2, 'evidence_driven': 0.21, 'contrarian': 0.21}}
5. Devil's Advocate Sonnet · legacy v2 Market price revealed
Reasoning Flaw Missing Info
Challenges
  1. Potential double-counting of settlement risk: sq1 already discounts heavily for settlement (60-65% settlement base rate), but the forecaster also notes adversarial discovery posture suggests NO imminent settlement. If OpenAI is fighting tooth and nail on discovery (preservation orders, contesting 20M logs), settlement probability conditional on this behavior should be lower than base rate. The 38% for sq1 may be too low given the specific adversarial signals.
  2. sq3 is described as 'subsumed in sq2' but the reasoning for sq2 only partially incorporates fair use. The verbatim reproduction evidence is genuinely strong against fair use under factor 3 (amount used) and factor 4 (market harm — NYT directly competes with ChatGPT for news consumption). The 55% conditional liability estimate may understate how much weaker OpenAI's fair use defense is for OUTPUT reproduction vs. training (which is what the N.D. Cal. ruling addressed).
  3. The combination 0.38 × 0.55 = 0.209 then 'rounded to 0.20 accounting for partial-ruling scenarios' is a downward adjustment that needs justification. Partial rulings could go EITHER way — a partial SJ ruling for NYT on one count would itself constitute 'NYT wins' under most resolution criteria. The rounding direction may be incorrect.
  4. Resolution criteria check: The question asks if a 'court finds' OpenAI infringed. Does an SJ ruling for NYT on liability count? Does a jury verdict count? Does a preliminary finding count? The forecaster doesn't explicitly address what counts as resolution YES, which matters because partial wins at SJ stage could resolve the market.
  5. The market at 69% with +13.1% over 30 days following the January 2026 discovery win is informational. While liquidity is low, the directional move tracks real case developments. The forecaster dismisses this as possibly 'NYT wins eventually' framing, but if the Kalshi market is specifically tied to the same Jan 2028 deadline, the gap deserves more scrutiny than 'mispricing.' Information gap: the exact Kalshi resolution criteria are not confirmed.
  6. Time awareness: 602 days until resolution with SJ briefing in April 2026 means a SJ ruling is plausible by late 2026/early 2027, leaving 12+ months for further proceedings. The forecaster's pessimism about timeline may be overcalibrated to historical cases (Authors Guild, Viacom) that didn't have aggressive SJ schedules already set.
Suggested adjustment: +8pp
Modest upward adjustment warranted for: (1) adversarial discovery posture should reduce settlement probability below the generic base rate used in sq1, (2) verbatim reproduction makes OpenAI's fair use defense materially weaker than a generic 50/50, and (3) partial SJ rulings should not be netted downward — a partial liability finding likely resolves YES. These are reasoning improvements, not market-chasing. The market at 69% is noted but adjustment is based on specific logic gaps.
Market comparison: Forecaster: 20.7%, Market: 69.1%. The 48-point divergence is large. Market has low volume/shallow orderbook (LOW-MEDIUM reliability), so it's not authoritative. However, the +13.1% 30-day move tracking the January 2026 discovery win suggests informed traders see the case strengthening. Some divergence is justified given market thinness and possible resolution-criteria ambiguity, but the forecaster's model may have stacked conservative assumptions (settlement base rate + slow timeline + 50/50 fair use) that compound multiplicatively. A forecast in the 25-35% range would better reflect the specific case posture without capitulating to a thin market.
6. Aggregator Opus · final calibrated forecast
30% (45% confidence)
I accept the Devil's Advocate's critiques on settlement (adversarial discovery posture genuinely reduces settlement probability below the 60-65% base rate for commercial copyright cases), fair use weakness for output reproduction (verbatim ChatGPT outputs of NYT articles are materially harder to defend than training-data fair use, and the N.D. Cal. Anthropic ruling addressed training, not output), and the partial-SJ-ruling asymmetry (a partial liability finding for NYT likely resolves YES, so rounding down was wrong). Adjusting sq1 to ~0.45 and sq2 to ~0.62 yields ~0.28, which I round to 0.30. I reject deferring to the 0.69 market price: liquidity is thin (107 contracts/day, shallow book), the 30-day run-up is noisy, and the gap between this market and base rates strongly suggests it's pricing 'NYT wins eventually' rather than the specific Jan 2028 deadline — a known framing error in thin prediction markets on legal questions with long tails. My 0.30 reflects the genuine strength of NYT's verbatim-reproduction evidence and active SJ schedule while preserving timeline risk that the market appears to ignore.