A problem for people who use AI
Making an AI prototype into something people can use
Creators and recipients of an image-editing setup, users of a government chat, and an author and playtesters of a personal game
Inputs and steps improved; continued effect unmeasured
Image editing was published October 21 with the implementation date not stated. The game is a July–October record. The government chat was published in August with the survey date not stated.
Editorial check: 2026-10-23
What people are paying attention to
In the October records of image editing and game making, adjustments were made to the situations where others use them and to the actual experience. A government AI chat also has design efforts.
What is hard
The issues were difficult input and leftover unwanted objects in image editing, and a first version that ran but was not fun in the game. The government chat needs to convey what it can do and how to check.
Workarounds being tried
Changes to input forms and steps, revision through actual play, and explanations of uses, primary-source checks and data use are being tried. Support for beginners' first question is a target for future improvement.
What the sources confirm
Each record reports changes to input, processing and explanation. However, no results compare recipients' continued use, time savings or improved experience under the same conditions.
What is still unclear
Whether things moved from a state the creator can run to a state recipients can keep using is insufficiently confirmed. Handling of transparent materials and similar, and beginners' question entry, still have issues.
Workarounds and results in more detail
Narrow the input and change the processing order
Reported creating and handing over
Built a form centered on Japanese and a photo. Tested eight photo conditions and changed the steps to cut out the product before changing the background.
Scope and limits: A record by one creator. Eight is the number of photo conditions, not users or a success rate.
Prepare the explanation before use and the path to primary sources
Done in design and tests
Explains uses, errors and primary-source checking, and data use. Supports primary-source links and accessibility.
Scope and limits: A report from GOV.UK Chat's provider. Number of survey participants, accuracy and effects of long-term use are not shown.
Support beginners entering their first question
Target for future improvement
No result of providing the support has been shown yet.
Scope and limits: The next challenge named by GOV.UK Chat's design lead.
Play it for real and revise
Done by the author
Revised through repeated play. After a report that the opponent complained of a headache in a sound-based versus game, they prepared a warning and another communication method.
Scope and limits: A personal game prototype. There is no comparative trial of fun, safety or development efficiency.
Read with care
- Image editing, the government chat and the personal game are efforts with different uses, creators and evaluation methods. They are not results measuring a common improvement effect.
- Image editing still needs visual checks and regeneration for transparent materials and Japanese text, and the number of users and retention are not given.
- The game's author is the same Simon Willison as in the voice-development card. They are not counted as two different users. There is also no independent verification of code quality or safety.
What to check next
A record of where users after handover got stuck, what they fixed, and how much they kept using it.
After an answer is generated, can the user check it and act?
In turning GOV.UK Chat into an app, they design what users looking for government information expect first and how they check answers. At the start of use it explains the uses and scope, the possibility of AI errors with source checking, and how conversation data is handled, and each answer leads on to the original guidance. On the small screen the explanation is condensed, and operation with enlarged text and with screen readers is also tested.
Some users get the impression that it can do far more than it can, and the provider is trying to set expectations that match its current functions. This is the design lead's report, not a source that proves with numbers a gain in understanding or a drop in inquiries. Editorially, it is material for observing the first question, checking the answer, and the action that follows.
Questions to think with
- Can a first-time user explain what they can hand over and where a person's judgment is needed?
- Can they go from the basis of an answer to the needed procedure or consultation? Where are the situations in which they cannot?
Editing that checked the primary sources. The questions to check in your own work are proposals for verification.