Microsoft Research Podcast: Jennifer Neville on AI’s surprising failures

Microsoft Research Podcast: Jennifer Neville on AI’s surprising failures

The Microsoft Research Podcast episode was published on October 6, 2026. Jennifer Neville argues that AI systems can fail in ways simple benchmarks miss, especially when users clarify requests over multiple turns or models work through long sequences of edits. In her team’s evaluations, performance that is high on single-turn tasks degrades significantly when the same task is specified across multiple turns; errors in longer workflows can also accumulate and cause further confusion. Neville says practical evaluation should reflect real user tasks, helping researchers find performance gaps and improve systems. She advises users to check answers rather than delegate everything: if a model gives a poor answer, trying again or asking differently may help. With oversight and verification, she says, current systems can be useful for productivity and speed. She also says users’ detailed feedback about what went wrong can help improve models, while noting that her team analyzes user data in privacy-preserving ways and does not fine-tune on users’ data or responses. Neville’s larger aim is for people and AI to work together in organizations. How much the current transformer architecture’s limitations can be addressed with wrappers, and what requires a different architecture, remains an open research question.

Publication: 0143f098-7706-1c3b-9958-057a43ebfbe7

attribution — pass: The reading correctly attributes the arguments to Jennifer Neville in a Microsoft Research Podcast conversation hosted by Chad Atalla. It keeps her findings, advice and privacy account as her statements rather than established fact.

Today, I’m joined by Jennifer Neville, a partner research manager at Microsoft Research

summary — uncertain: The summary is mostly faithful and readable, and transformer architecture is defined in the term library. However, the closing paragraph and the question both depend on 'wrappers', which is neither explained nor defined. In the source, wrappers are the agentic harnesses and back-end reasoning built around the model, and a lay reader cannot follow the open question without that. Smallest repair: add a plain gloss, such as 'wrappers (the agentic harnesses and reasoning layers built around a model)', or add 'wrapper' to the term library.

A lot of those limitations can be solved by the wrapper around the models—the agentic harnesses that we have around the models right now—and the reasoning that goes on, on the back end.

date — pass: The reading's October 6, 2026 date matches the UTC calendar day of originalPublished. The page is a publisher page, not an archive copy, so no archive disclosure is needed. The reading uses no post-publication passages and adds no hindsight framing. The page was modified on October 20, 2026, but every relied-on passage is plainly part of the original episode transcript.

support — pass: Each claim appears verbatim in the transcript: single-turn performance degrading over multiple turns, errors accumulating in long workflows, the advice to check answers and avoid full delegation, the value of detailed feedback, the privacy account and not fine-tuning on user data, and the open question about transformers. The cited passages are plainly part of the original transcript rather than a later edit. The reading adds no unsupported benefits or costs.

So just to be clear, we’re not fine-tuning on your data or your responses

media — pass: No media was selected and no verified media candidates were supplied, so there is no media identity or relevance issue.