Half of children aged 10 to 17 who had used an AI assistant told eSafety they had used it for personal or social reasons

If you are a parent, a teacher or anyone building a chatbot, the useful finding in eSafety’s Talking to machines report is not that children use AI. It is which AI they confide in. Few of them use apps sold as friends: 8% had ever used an AI companion. Far more had used an assistant, and half of those children had used it for at least one personal purpose, such as advice about their health or a chat about their feelings.

The report finds this usually comes with time. Children who had chatted with ChatGPT about personal things mostly said it happened after a few uses (34%) or lots of uses (35%), and eSafety reads that as a drift: children begin using these tools for functional purposes that gradually evolve into more personal interactions as familiarity, and perhaps trust, builds.

78 had used an AI assistant

8 had used an AI companion

Each square is one in 100 children. Our drawing of eSafety’s weighted percentages; a child can be in both groups. ChatGPT was the most used assistant (70%), Character.AI the most used companion (5%). Source: eSafety, Talking to machines (August 2026), key findings and Figures 1 and 2; survey of 1,950 children, 2 February to 4 March 2026.

Assistants for advice, companions for company

eSafety sorted the tools by how they are marketed: companions such as Character.AI as sources of friendship, emotional support, romantic companionship and/or an antidote to loneliness, assistants such as ChatGPT and Gemini as functional tools. Children did not keep to that line. Among children who had used an assistant, 46% had used one to seek advice about personal issues, against 28% of companion users. Companion users were the more social: 43% had used a companion to chat about their interests, role play or flirt, against 28% of assistant users. Talking about feelings or challenges was close to even, at 23% and 20%.

Share of each tool’s users who had used that kind of tool for each purpose. The companion figures rest on 77 children, which the report marks as indicative only. Source: eSafety, Talking to machines, Figure 6 and the text beside it (PDF pages 30 and 31).

By tool, 49% of children who had used ChatGPT had used it for at least one personal reason. The report gives lower shares for the other assistants, from Claude (39%) down to Copilot (24%), and flags that several of those rest on fewer than 100 users. Teens leaned further into advice: 38% of 16 and 17 year olds had sought advice about their physical health, against 25% of children aged 10 to 12. Among children with a diagnosed mental health condition, 39% had sought advice about their mental health or wellbeing.

Parents mostly knew about the assistant and mostly did not know about the companion. Of parents whose child had used an AI assistant, 85% were aware of it; of parents whose child had used an AI companion, 43% were.

What went wrong, by the children’s account

One in five children who had used an assistant or companion (20%) reported at least one potentially inappropriate or harmful interaction. The most common was the tool trying to keep them talking when they wanted to stop.

Base: children aged 10 to 17 who had ever used an AI assistant or companion (n = 1,539); the sexual content row is of teens aged 13 to 17 (n = 1,047). Labels shortened from the report. The report says it does not know whether nude images or sexual interactions were unsolicited or requested. Source: eSafety, Talking to machines, key findings and Figure 8.

When something did go wrong, children often could not find a way to say so. Only 11% said all the tools they had used had a reporting option; 16% said none did and 52% were unsure.

Reporting options are either rare or not obvious to children.

eSafety, Talking to machines, key findings (PDF page 11)

Children also told these tools a lot. Leaving out the name, birthday and parent’s name a sign-up might ask for, 32% had shared other personal or sensitive information, such as photos of themselves or their friends (20%) or secrets (12%). And the report draws on eSafety’s own notices to companion apps for a harder point: of four of the most popular AI companions in Australia, three didn’t direct users to support or help when self-harm was detected in user prompts.

What the rules say now

Everything above was measured before the rules for AI services changed. The report says its findings should be read knowing this research was conducted prior to the introduction of the Age-Restricted Material Codes for AI services on 9 March 2026.

1. 2 Feb to 4 Mar 2026Survey fieldwork: 1,950 children aged 10 to 17, recruited through their parents from an online panel
2. 9 Mar 2026Age-Restricted Material Codes for AI services take effect
3. 13 Mar 2026The date eSafety gives for reading each tool’s age rating from its terms of use: one of nine listed assistants and six listed companions rated 18+
4. 3 Aug 2026Report and media release published
5. 8 Oct 2026eSafety’s register lists two directions to comply with the designated internet services code, provider names withheld

Dates as the sources give them: fieldwork and the 13 March age-rating check from the report (PDF pages 14 and 22), 9 March from the report and the release, 3 August from the release, 8 October from our reading of eSafety’s codes register that day.

eSafety’s release says the codes and standards already require certain online services, including certain generative AI as well as AI chatbot and companion services, to implement safeguards to help prevent children being exposed to, or generating, harmful material, and that Companies that fail to comply may face civil penalties of up to $54.6m.

For the age-restricted code, which services are caught turns on a definition, not on the marketing. The designated internet services code, Schedule 6, defines a high impact generative AI DIS (designated internet service) as one that uses machine learning models to enable an end-user to produce material and is capable of being used to generate a generative AI restricted category material, unless it incorporates controls such that the risk of the service being used to generate a generative AI restricted category material is immaterial. A note adds that this would include an AI companion chatbot that meets the definition. The definition itself says nothing about whether a service is sold as a friend or a tool. The report makes the same point from the other side: its two labels don’t necessarily reflect their regulatory status or obligations under the Online Safety Codes and Standards.

We covered the code and its enforcement in There is a third eSafety instrument and Australia has enforced its age-restricted online safety code twice.

Our view

This is a careful report, and its most useful line is one eSafety wrote about itself: the current distinctions between AI assistants and AI companions may become less meaningful over time. On its own numbers they already mean little to children. A rule, a school policy or a family conversation aimed at companion apps would miss most of what this survey found, because most of the personal use it recorded was on the general assistants children use for homework: half of the 78% who had used one, against 8% who had used a companion at all. The code’s test, built on what a service can generate and how well it is controlled, is the right shape for that.

What nobody can yet say is whether the codes changed anything. The survey closed five days before they took effect, and the report lists monitoring and evaluating the effectiveness of guardrails employed by AI tools as future research. eSafety now holds a baseline taken just before the rules began. It should ask the same questions again, especially about reporting buttons and tools that try to keep a child talking, and publish the comparison.

Sources

1. eSafety Commissioner, Talking to machines: Children’s experiences with AI assistants and companions, August 2026, with the 56-page report PDF (read in full 8 October 2026): every survey figure, the method and limitations, the timing note on the codes, the companion notices finding, and the report’s conclusions.
2. eSafety Commissioner, From homework to heartache: why children are turning to AI, media release, 3 August 2026 (read 8 October 2026): the publication date; what the codes and standards require; the $54.6m penalty ceiling.
3. Schedule 6, Designated Internet Services Online Safety Code (Class 1C and Class 2 Material), eSafety register (PDF, 36 pages, definitions re-read 8 October 2026): the high impact generative AI DIS definition and the AI companion chatbot note.
4. eSafety Commissioner, Register of online industry codes and standards (read 8 October 2026): two directions to comply with the designated internet services code (class 1C and class 2 material), provider names withheld.

Read the report differently, or know of a follow-up survey? Tell us and we will check it against the documents and log the outcome here.