Education

OpenAI plans a teen college planner as safety testing raises concerns

By Saint Zero Day · 2026-10-08 · 4 min read

College Planner is forthcoming, with benefits still untested. Independent testing of the surrounding teen experience reports safety gaps, with important methodological limits.

OpenAI announced on October 7 that it plans to add College Planner to ChatGPT for Teens, organizing application requirements, deadlines and financial-aid tasks. The initial audience is US students in grades 10–12 considering four-year colleges. The announcement describes a forthcoming service, without an exact launch date, pricing details or an evaluation showing improved applications or enrollment. [1]

The company also announced support for College Advising Corps and Boston Children’s Hospital’s Digital Wellness Lab, alongside additional study tools. Its stated ambition is to make educational help more accessible. OpenAI reports nearly 1.2 million teens used Learning Visualizations in one week and more than 180,000 used Study Mode. Those figures measure use; they do not establish learning gains or successful college transitions. [1]

There is a concrete rationale for organizing administrative work. Education researchers Katharine Meyer and Lindsay Page describe how timely information can help students bridge the gap between intending to attend college and completing necessary steps. At Georgia State, an earlier chatbot intervention reduced summer melt by about three percentage points. That system combined reminders, responses and staff assistance, making it relevant to College Planner’s proposed task support, but not an evaluation of this new product. [2]

Why it matters

The comparison also carries a warning. The researchers describe a national FAFSA reminder experiment involving about 10,000 students that produced no overall improvement in filing, aid receipt or persistence. Successful local programs benefited from familiar senders and targeted information. Editorial inference: a convenient planning interface may help with confusion, but its reach alone cannot establish effectiveness, and reminders cannot remove structural barriers such as insufficient money. [2]

Human advising could matter particularly for students with limited access to individualized guidance. College Advising Corps describes a model that places recent graduates in under-resourced high schools through local partnerships. Its advisers support several possible destinations, including community college, trade school and military service. That breadth is relevant: useful guidance must help students consider different paths, while the announced planner’s initial scope is narrower. The nonprofit’s existing work does not demonstrate the effectiveness of its forthcoming AI fellowship. [6] [1]

Learning support poses a different question: does assistance leave students more capable afterward? A peer-reviewed, Pennsylvania-led trial involving nearly 1,000 high-school mathematics students in Turkey compared two GPT-4 tutors with no AI access. A basic tutor improved assisted practice performance by 48%, but students subsequently scored 17% worse when working without it. A tutor supplied with teacher-designed guidance increased practice performance by 127% and largely avoided that penalty. [3]

The guarded tutor did not produce a positive effect on the unaided exam. It used labor-intensive, problem-specific solutions and hints, rather than relying solely on a general instruction to teach. This older trial concerns a different model, subject and setting; it neither replicates nor disproves OpenAI’s new tools. It does show why assisted output and retained understanding need separate measurements. The paper’s later correction updated an author affiliation. [3]

More immediate adverse evidence concerns the surrounding teen experience. Common Sense Media’s Youth AI Safety Institute tested more than 4,000 prompts before and after its August launch. It reported no parental alerts in a dedicated experiment across more than a dozen fresh linked accounts discussing crises, while broader testing produced four alerts. Its finding is therefore unreliable notifications under tested conditions, rather than proof that notifications never work or a measured rate of harm among actual teenagers. [7]

The assessment includes a material dispute: OpenAI said notifications require approximately three hours after account linking to activate. The institute acknowledged some tested accounts were linked for less time, but said others exceeded that window. Its sequential testing also confounds underlying model changes with teen settings, excludes several modalities and lacks calculated inter-rater agreement. OpenAI reviewed a draft for factual accuracy without editorial control. These are external product tests, not an unrestricted audit. [7]

The institute also reported easily bypassed study controls and working refusals of sexual roleplay. Axios reports OpenAI considers leaving Study Mode intentional flexibility. The dispute partly concerns what a safeguard promises. Common Sense discloses funding from the OpenAI Foundation while retaining editorial independence; its judgment should be assessed through its evidence and limitations, rather than treated as financially detached confirmation. [7] [8] [9]

Youth participation offers another route to improvement. The Digital Wellness Lab’s council is entering its fifth year, with two tracks running from October 2026 through May 2027. Its Architects track, for ages 16–20, includes hands-on testing and published industry recommendations. This is structured consultation with a selected group, not a representative survey of teenagers. Recommendations and documented changes could make its influence assessable. [4]

The Lab says industry donations support some work, but it rejects research for hire and does not endorse particular products. That distinction matters when a partnership appears alongside a product announcement. Editorial assessment: the planned college tool has a plausible practical purpose, while benefit remains unverified and surrounding protections face specific challenges. Comparative application outcomes, unaided learning tests, reproducible safety retesting and public records of responses to youth recommendations would provide stronger grounds for judging the initiative. [2] [3] [4] [5] [7]