Making AI work for online students: What the OfS AI report means for online and partnership delivery

Today's Office for Students and Advance HE report on AI does not mention online or partnership provision. Many online programmes already use authentic assessment, but the report is a useful prompt to review where online assessment strategies meet its expectations and where they could go further.

This morning the Office for Students (OfS) published Making AI work for students: Reflections on institutional practice, a joint project with Advance HE, together with new guidance on how AI fits within its existing quality conditions (OfS, 2026a, reference, 2026b, reference). The research drew on more than 300 students and staff from over 120 institutions, gathered through round tables, workshops and interviews between June and September this year.

The effect of generative AI on assessment has been debated across the sector since ChatGPT was released in late 2022, and earlier this week I argued that AI-proofing traditional assessments misses a critical part of the challenge: unsupervised coursework cannot be made reliably resistant to AI, and adding safeguards to every task increases the burden on students who already have the least time to absorb it. In particular, online postgraduate taught (PGT) programmes should deliver fewer, more secure, intentionally designed assessment points instead, and the OfS report reaches a similar place by a regulatory route. The report is aimed mainly at campus-based delivery, but with the increasing pace of online development across the sector, its implications for online provision need to be considered to ensure future compliance and fairness for students. Neither the report nor the guidance mentions online, distance or partnership delivery specifically, though they raise some distinct questions of their own.

What the OfS is asking for

The core message sits under Condition B4, which requires institutions to make sure each assessment is valid and reliable, meaning that it does result in students demonstrating the knowledge and skills being assessed. The OfS press release adds that "in many cases, this will require the use of secure (AI-free) assessments" (OfS, 2026c, reference). The guidance defines secure assessments as those designed to prevent the use of AI, and lists in-person examinations, presentations, vivas, performance-based assessments and practical tasks as examples (OfS, 2026b, reference).

The guidance also covers Condition B1, on keeping courses up to date with the skills students need, and Condition B2, on supporting students, including clear policies on permitted use and fair access to paid tools. The press release restates that the usual sanctions remain available: specific conditions of registration, monetary penalties, and suspension of access to student support or OfS grant funding. Josh Fleming, the OfS Director of Strategy and Delivery, describes AI as a generational risk to the system as it currently stands (OfS, 2026c, reference).

None of this introduces a new condition, because participants in the research told the OfS that AI-specific conditions could be counterproductive and asked instead for clarity on how the existing ones apply. The new guidance responds to that request by setting out how the current conditions should be interpreted across the sector.

Why online is different

The report found that it is frequently possible to complete unsupervised assessments to a high standard using AI tools, without the student developing the intended skills. Students described peers who used large language models, sometimes for only a few minutes, and received grades similar to those who did the work themselves. This evidence is anecdotal rather than statistical, but it matches the national survey picture: 94% of full-time undergraduates surveyed by the Higher Education Policy Institute (HEPI) say they use generative AI to help with assessed work (Stephenson & Armstrong, 2026, reference).

On campus, unsupervised coursework sits alongside examinations, labs and seminars, whereas on some online programmes unsupervised coursework makes up much of the assessment, even where it is designed to be authentic and work-based. Our students are spread across time zones, study asynchronously and submit through a virtual learning environment, and our programmes are designed to recruit at scale. Each of those features is a reason people choose online study, and each needs to be considered in light of the gap the report describes.

The report also closes off the easy answer, finding that AI detection tools do not consistently and reliably identify AI use, and the guidance warns against giving their outputs disproportionate weight (OfS, 2026b, reference), which is consistent with the research I reviewed in my earlier article (Weber-Wulff et al., 2023, reference). Online providers are unlikely to resolve this with software that checks submissions after the event; the answer has to come from improved and consistent assessment design.

What it means for partnerships

In my role as Director of Academic Practice, I work on the delivery partner side of online partnerships with UK universities. In these models the registered provider awards the degree and answers to the OfS, while the delivery partner often authors the content, supports students and sometimes manages marking workflows. Regulatory responsibility stays with the university, while much of the practice the regulator is now looking at sits with the partner.

Following the publication of the report, I see three key challenges for the future of online delivery:

• Assessment design. Where assessment is co-designed, or templated across a portfolio of programmes, the partner's design decisions directly shape whether the university can show that its assessments are valid and reliable under B4.
• Feedback. The guidance expects oversight wherever AI is used to prepare feedback, so that it is accurate and aligned to the assessment criteria (OfS, 2026b, reference). Any AI-assisted marking or feedback tool in a shared workflow needs a clear line of accountability, with human review and a record of how it was used.
• Learning materials. Some students in the research felt that staff use of AI had reduced the quality of learning materials, and content produced at scale for online delivery could be read in that light. Research on how readers judge whether text is AI-generated has found that they rely on unreliable cues about style and phrasing and cannot reliably tell the difference (Jakesch et al., 2023, reference), so polished, consistent materials may be assumed to be AI-produced whether or not they are.

I expect universities to start asking their delivery partners for this evidence before the regulator asks them, and partners who can already show an appropriate assessment strategy will be well placed to respond.

What good looks like online

The report describes three approaches that participants were already using to keep assessment valid (OfS, 2026a, reference), and all three can be delivered without bringing online students onto a campus. Building on the AI assessment framework I set out in a previous blog, I would suggest how each translates to online programmes:

1. Secure assessments. The report defines these as assessments designed to prevent students from using AI while completing them, including in-person examinations, presentations, vivas, short discussions that test students' understanding of work they have submitted, and practical or performance-based tasks. Live online vivas, recorded presentations followed by questions, and short timed tasks all work in an online setting, and QAA (2023), reference has already described a retreat to handwritten examinations as regressive. Secure points do not need to be in every module, but there should be enough of them, at meaningful points in the programme, to stand behind each award.
2. Process-based assessments. These assess how students produce their work alongside the final output, and some participants described checking skills such as completing a literature review or summarising a text at several checkpoints, or "stage gates", during a module. Online programmes are well suited to this, because drafts, decision logs and reflective portfolios already pass through the virtual learning environment and can be paired with a short viva for a sample of submissions. This needs care, however, because added to every task it becomes the over-assessment problem in another form.
3. Assessments where AI use is permitted or required. Examples in the report include critiquing and interrogating AI-generated output, or using AI to produce a first draft that students then edit and improve, and these tasks can also be used to evaluate students' AI literacy. This is the open lane of my earlier framework, and it matters because the report found that students see AI skills as important for their careers, while staff AI literacy training is frequently optional and often has low take-up. Both staff and student AI literacy therefore belong in programme design rather than in optional extras, and our Academic Study Skills project has shown me how much online PGT students value support that is built around their time constraints.

Whichever mix a programme uses, students need to know which rules apply to which task. Students in the research were frequently unclear whether and how AI was permitted, and a single plain-language statement, agreed between the university and its partner and repeated consistently in every module, would remove much of that uncertainty.

Corbin et al. (2025), reference would describe the three approaches as structural changes and the policy statement as a discursive one. Discursive changes matter for students' confidence, but only structural changes give the university evidence it can put in front of the regulator.

The opportunity

Writing on Wonkhe, Debbie McVitty (2026), reference observes that the report's expectations read as both a baseline and wildly aspirational, and that the harder question is how institutions will manage the transition. I agree that structural change is the hard part; according to the report, relatively few of the institutions that contributed to the research have yet adopted a strategic approach to AI.

Online and partnership teams do have one structural advantage: because we design programmes centrally and update them systematically, a change that might take a campus faculty several years can reach a whole online portfolio within a few delivery cycles. The same is true of professional accreditation, where I have seen online programmes evidence equivalence of outcomes through deliberate, portfolio-level choices rather than module-by-module fixes.

In a forthcoming article I will turn to the question both this report and my last article leave open: if we have fewer, more secure assessment points, what do we want graduates to be able to show at them?

References