How WhimsyCat Meets the DfE AI Safety Standards Without a Chat Box

When the Department for Education published its updated Generative AI product safety standards on 19 January 2026, it set out what an AI tool must do to be safe in a school: detect when a learner is in distress, avoid manipulating them, filter harmful content, guard against dependence, and never train on a child's work without consent (DfE, 2026). Read them closely and a pattern emerges. Almost every requirement is about the risks of letting a generative chatbot hold a conversation with a child.

WhimsyCat, our AI science tutor, takes a different path. It does not give students a chat window at all. A pupil never types a prompt and never receives generated text back. Everything WhimsyCat knows about a learner, it infers from their actions in the virtual lab: which equipment they pick up, the order they work in, where they hesitate, and what they get wrong. That single design decision means most of these standards are met by architecture rather than bolted on as filters. They also arrived alongside the government's plan to put AI tutoring in front of up to 450,000 disadvantaged pupils (DfE, 2026), so the question of what "safe" means is no longer abstract.

The Standards Assume a Child Talking to a Bot

Look at what each requirement takes for granted. The distress requirement assumes a child might type something worrying that the AI must notice. The manipulation and dependence requirements assume an ongoing conversation a child could be drawn into. The content-filtering requirement assumes the AI generates free text that could turn harmful. The data requirement assumes the AI is collecting what the child writes. Every one of these risks flows from the same source: an open generative channel pointed at a child. The DfE is right to regulate it, because that channel is how most AI tools work. The department has been candid that the evidence these tools improve outcomes is still limited (Tes, 2026), which is all the more reason to start from safety.

WhimsyCat Has No Chat Box, and That Removes the Risk Surface

Students do not converse with WhimsyCat. They do experiments, and WhimsyCat watches. Because there is no text box, there is nothing to jailbreak, no open generation to filter, and no conversation to be manipulated by or to grow dependent on. The harm surface that the standards are written to contain is simply not present in our product. This is not a claim that we filter those risks more cleverly than other tools. It is that the channel through which those risks occur does not exist here, so there is far less to filter in the first place.

It Cannot Simply Hand Over the Answer

The standards ask developers to mitigate cognitive deskilling and to report to teachers on how much cognitive offloading a tool is delivering. With no "tell me the answer" chat, a pupil cannot offload by asking. When WhimsyCat infers from a pupil's actions that they are struggling, it responds with guidance pitched to keep them working rather than with the solution. Our process-based assessment records how each pupil approached the task and surfaces that to the teacher, which is exactly the visibility into cognitive offloading the standard asks for, produced as a by-product of how the tutor works rather than as a separate reporting feature.

Distress and Emotional State, Inferred From Behaviour

WhimsyCat detects a pupil's state from what they do, not from parsing what they write. Repeated failed attempts, long hesitation, or abandoning a task are signals of frustration or disengagement, and WhimsyCat surfaces them to the teacher, who remains the human in the loop. We are careful about the line here: WhimsyCat is not a counsellor and does not replace a school's safeguarding processes. But the specific risk the distress standard targets, a child confiding something serious to an AI in conversation, does not arise when there is no conversation to confide in. The teacher stays the point of human contact, and we give them better information to act on.

Manipulation, Dependence, and Content, Designed Out

No chat means no coaxing the tutor into saying something it should not, and no parasocial relationship with a bot that presents itself as a friend. A pupil's attention is on the experiment, not on a conversation that could deepen into reliance. There is no open text generation that might surface harmful or age-inappropriate material, because WhimsyCat does not generate free text to pupils at all. The standards treat these as risks to be mitigated. For us they are consequences of not building an open chatbot in the first place.

Pupil Data Is Actions, Not Conversations

The signal WhimsyCat works from is interaction data inside the lab, not a transcript of everything a child has typed. We collect only what is pedagogically necessary and never train models on pupil work without consent. In a period when education data breaches have made every parent and data protection lead more cautious, an AI tutor whose data is lab actions rather than free-text conversations is a smaller and clearer thing to govern, and a smaller thing to lose.

Meeting the Standards by Architecture

The DfE standards are a useful test, and many tools will now race to retrofit distress detection, content filters, and offloading reports onto a chatbot that was never designed for a classroom. WhimsyCat does not need to retrofit them, because it was built as an action-observing tutor inside a science task rather than a conversation pointed at children. For a school doing due diligence, the fastest question to ask is the one the standards imply but do not say outright: does the student get an open chat window with the AI? If the answer is yes, every standard becomes a live risk to manage. If the answer is no, most of them are handled before you start. We built WhimsyCat so the answer is no.

Related Articles

• Emotional Intelligence in AI Tutors: How WhimsyCat Detects Frustration
• Why Purpose-Built AI for Education Makes the Difference
• Ten Questions to Ask EdTech Vendors After PowerSchool
• AI Science Tutors: What Actually Works in the Classroom