cd /news/artificial-intelligence/mit-researchers-show-wrong-ai-thresh… · home › topics › artificial-intelligence › article
[ARTICLE · art-172270] src=alphasignal.ai ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

MIT Researchers Show Wrong AI Thresholds Quadruple Insurance Tail Risk

A revised MIT/Oxford paper by Dimitris Bertsimas and Agni Orfanoudaki finds that tuning binary AI classifiers for maximum accuracy can carry up to four times the worst-case (CVaR) loss of risk-aware thresholds, with threshold gaps of 0.159 in screening and 0.591 in diagnostic mammography producing tail-risk penalties of 77% and 297%. The authors propose insurance contracts that make risk-aware thresholds a condition of coverage, under which well-calibrated firms gain 7-8% and poorly calibrated firms up to 105%, while partially aligned firms can lose under strict mandates. The work connects classifier design choices to liability exposure as commercial AI insurance products such as Armilla AI's policy (limits up to US$25 million) and Munich Re's aiSure already reach the market.

recommendation

Evaluate binary classifiers on CVaR at a 95% confidence level rather than validation accuracy alone, since accuracy-maximizing cutoffs carried up to four times the worst-case loss in the mammography study.

more in playbook: Evals and testing for LLM apps
read3 min views2 publishedNov 19, 2026
MIT Researchers Show Wrong AI Thresholds Quadruple Insurance Tail Risk
Image: Alphasignal (auto-discovered)

A revised MIT/Oxford paper shows that tuning AI classifiers for accuracy can quadruple worst-case liability, and proposes insurance contracts that force safer decision thresholds.

  • Bertsimas and Orfanoudaki's revised Algorithmic Insurance paper links classifier thresholds directly to tail liability risk.
  • In their mammography study, accuracy-maximizing thresholds carry up to four times the worst-case (CVaR) loss of risk-aware thresholds.
  • Threshold gaps of 0.159 (screening) and 0.591 (diagnostic) produce tail-risk penalties of 77% and 297%.
  • The proposed contracts require risk-aware thresholds: well-calibrated firms gain 7-8%, poorly calibrated firms up to 105%.
  • Partially aligned firms can lose under strict threshold mandates, so the authors call for flexible contract design.
  • Extensions cover model drift, renewal timing and human oversight. The method is limited to binary classifiers, and no code release is confirmed for this revision.

Every deployed binary classifier needs a cutoff above which it returns a positive result. Teams often choose 0.5 or the threshold that maximizes validation accuracy. In the revised Algorithmic Insurance paper, Dimitris Bertsimas and Agni Orfanoudaki report that this convention can sharply increase financial exposure. In their mammography case study, accuracy-maximizing thresholds produce up to four times the tail risk of thresholds optimized for conditional value-at-risk, or CVaR. The authors also propose insurance contracts that make the safer threshold a condition of coverage.

One flaw, many claims #

Insurance pricing relies on pooling losses whose frequency, severity, and correlations can be estimated. A shared algorithmic flaw complicates that calculation because one defect can affect thousands of patients, drivers, or customers across different settings. The resulting claims may arrive together while varying widely in size, leaving insurers with limited data for pricing and developers without a clear link between system design and liability exposure.

Commercial coverage has already reached the market through products such as Armilla AI’s standalone affirmative AI liability policy, with limits up to US$25 million, and Munich Re’s aiSure suite, which covers contractual liabilities, financial losses, and legal liabilities arising from AI performance errors. An agentic insurance survey describes additional structures, including additions to errors-and-omissions policies and standalone coverage. Pricing remains largely bespoke, and the paper offers a mathematical framework for connecting premiums to a deployment setting that developers control.

The tail changes the cutoff #

At a 95% confidence level, value-at-risk, or VaR, is the loss threshold exceeded in the worst 5% of cases. CVaR averages the losses beyond that threshold, measuring the expected severity of those worst outcomes. The authors’ earlier presentation slides illustrate the same definition. Insurers use this measure because extreme losses drive capital requirements and solvency exposure.

Mammography shows why the distinction affects model behavior. A false positive can trigger an additional clinical work-up, while a false negative can delay a cancer diagnosis and lead to a much larger malpractice claim. Changing the classifier threshold shifts the frequency of both errors. The study selects the threshold that minimizes CVaR across their loss distributions.

This story is for Pro members #

You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @mit 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/mit-researchers-show…] indexed:0 read:3min 2026-11-19 · —