The introduction uses critical-supply logistics rather than school trips, shows an interactive route-cancellation example, and connects documented AI-error cases to the research questions. Afterward, visit the experiments page to see the replayable V0–V5 graphics.
A visual introduction · Start here
Why an artificial intelligence system should remember not just its answers, but also the reasons behind them.
This page starts with a practical medical-supply routing example. No programming background is needed.
01 · A familiar problem
Imagine a hospital asking an AI-assisted logistics system whether essential medical supplies can arrive today. It answers yes because the shipment can travel by highway or rail. What happens when the highway closes?
Cancel one transport option, then the other. Watch how the answer changes. You can bring either option back at any time.
Both routes point toward the same answer. Either one can support it.
Both routes are available.
Either the highway or the rail route could make the delivery possible. A reason is information that helps justify an answer.
The highway is no longer available. But the rail route still supports the answer. The assistant should not simply forget that alternative.
Now neither travel option is available. The earlier answer needs to be reconsidered.
02 · Why this research question matters
AI assistants can make mistakes for many reasons. One possible failure is carrying an earlier answer into a new situation without checking whether its supporting information is still reliable or relevant. An answer can sound confident while relying on an outdated assumption.
Imagine someone asks an AI assistant which road is available during an evacuation. The assistant has an earlier report that a bridge is open. Later, that bridge closes.
The assistant should recognize that the older recommendation has lost its support, reconsider affected advice, and avoid claiming it has a safe route without sufficient current evidence.
This is a hypothetical illustration of a possible AI failure—not a reported real incident and not a scenario tested in CRSE. The early experiments below do not prove safety in emergency or other real-world decisions.
These real incidents show why errors involving sources, authority and suitability to context deserve attention. They do not demonstrate that CRSE could have prevented the incidents or identify their exact internal technical causes.
In a 2024 tribunal decision, an airline chatbot had given incorrect advice about claiming a bereavement fare after travel. A passenger relied on it and Air Canada was held responsible.
Research question: Can answers be re-evaluated when their supporting information is wrong or conflicts with a better-authorized source?
Source: American Bar Association, 2024 ↗In 2024, investigators documented New York City's MyCity chatbot giving incorrect answers to business owners about tenant protections and worker rights, even though it appeared to convey official guidance.
Research question: How should systems distinguish an answer that sounds authoritative from a conclusion actually supported by the governing source?
Source: The Markup / THE CITY, 2024 ↗In 2023, the National Eating Disorders Association paused its Tessa chatbot after reports that it suggested weight-loss practices to people seeking eating-disorder support.
Research question: Can a system withhold an otherwise familiar answer when the situation makes that advice inappropriate?
Source: WIRED, 2023 ↗An error in an upstream assumption can affect many later decisions. Our V2 experiment addresses this in twelve deliberately connected synthetic examples; no comparable real-world operational system was tested. See V2 →
These questions motivate the experiments. The existing tests explore narrow, deliberately constructed examples—not every way that AI can make errors.
03 · The idea
Evidence is information used to support a conclusion. A conclusion is an answer or determination. The arrow means “helps support”; it is not a movement animation and does not tell you whether something is currently valid. Colored boxes show the status.
A correction changes what information can still be relied upon. When a reason is no longer supported, a conclusion that relied on it may need to change too.
Context means the circumstances in which an answer applies. A timetable can be correct for Monday without being suitable for Tuesday. A stored fact may remain correct within its original circumstances without being usable everywhere.
03 · What has actually been tested
The critical-supplies examples are teaching illustrations, not the original data used by the experiments. The first four bounded tests used made-up test cases (called synthetic cases) and investigated: withdrawing evidence, rejecting an unauthorized change, updating twelve explicitly linked dependent conclusions, and restricting a conclusion to the context where it applies.
These experiments do not establish that the system independently understands natural-language meanings, works at large scale or outperforms other approaches.