Transcribing for Fun and Profit

We are firm in our dislike of all things related to Large Language Models. In our vaunted opinion, such systems are the stuff of idiocy and nonsense, and their ability to appear human-like in their responses is, at best, deeply suspicious, and horrifically misleading at worst. If only more people shared our view, perhaps there would not be a constant stream of resources being sacrificed to the addled machine god.

Our latest concern, hitting somewhat close to home, is the use of a system known as Genio in the University of British Columbia here in Vancouver. Genio is a note-taking and transcription service, and it is now being used in place of actual human note-takers for students who require accessibility accommodations in lectures.

Now, this in and of itself is somewhat less of an issue than usual. The use of computer transcription is at least well known, and has been used in office and academic contexts for decades. Note that the original and probably the most well-known voice-to-text program, Dragon Naturally Speaking, was first released in 1997. Genio itself began as an audio transcription tool back in 2005, and its purpose was accessibility for students. The technology of automated transcription is, in fact, a sensible use for a neural network model trained on voice data; the model’s purpose, in this case, is to ingest audio data, turn it into tokens, and match up those tokens to text words which are then reassembled and sent to the output.

We have mentioned the issue of hallucinations before, and you may well ask how does this apply to this kind of model. Hallucinations happen regardless of the type and size of model, after all. But hallucinations in a model that is effectively decoding audio is far less of a problem than those coming from an LLM. There is far more “wiggle room”, as it were, for the output to be a little off without the actual meaning being lost or altered. This is not to say that it can be perfect (it is still a probabilistic system) but such a model is not going to, say, suggest that glue on pizza is acceptable.

So, let us turn to UBC and its decision to force all students requiring note-taking accommodations to use Genio. One student in particular was kind enough to provide a copy of the email below to us.

The transcription is as follows:

Hello,

We are excited to share a major update regarding note-taking accommodations through the Centre for Accessibility.

Beginning in September 2026, students approved for note-taking support accommodations (previously provided through peer note-taking or audio-recording accommodations) will instead receive access to Genio Notes, a purpose-built note-taking platform designed specifically for post-secondary learning environments.

Genio Notes enables you to record lectures, generate transcripts, organize course materials, and create personalized notes—all within a single platform. The tool has been reviewed and approved through UBC's privacy and security processes and provides a consistent, accessible, and effective approach to note-taking support.

With Genio Notes, you can:

-Record lectures and create searchable transcripts
-Add labels, comments, and notes while listening
-Import lecture slides and other course materials directly into your notes
-Organize and review content in one place

Many students find that Genio offers greater flexibility and independence by allowing them to customize their notes according to their own learning needs and preferences. The platform is already being used for this purpose at a growing number of post-secondary institutions in Canada and internationally.

To help you prepare for September, your Genio account will be available during the summer so that you can explore the platform and become familiar with its features before classes begin.
More information, including account activation instructions and training resources, will be shared shortly.

We have already shared this update with Faculty and academic leadership and will continue to update instructors.

In the meantime, we encourage you to review our Frequently Asked Questions on our CfA Canvas course here: Module 1.7. If you have not already enrolled in the Canvas course, you can do so here: CfA Canvas Course Enrollment

Sincerely,

The Centre for Accessibility

Genio’s main purpose, that of transcription, uses machine learning models provided by AssemblyAI and Deepgram. Despite the name, these models are most certainly not “AI” in the same way that a Large Language Model is understood to be; they are far more pedestrian and certainly more useful, as we discussed above. Genio also includes several features that are most definitely in the realm of AI, as the functionality depends on OpenAI’s LLMs. These features are specifically Quiz Me, Study Notes, and AI Outline, all of which operate on the transcriptions produced by AssemblyAI/Deepgram, and involve sending prompts and data to OpenAI.

We have confirmed, through contact with Genio and through our own research, that student data is restricted to the United Kingdom, where Genio is based. We have also confirmed that the AI features are most likely processed through Microsoft’s Azure data centre in the south of London. Genio, to their credit, have made an effort to be transparent about where the data is kept, with the caveat that both Microsoft Azure and OpenAI retain user data for 30 days for the purpose of preventing abuse before it is deleted.

Unfortunately, we trust Microsoft and OpenAI about as far as we can hurl our laptop across an open field and, to be clear, we are not athletically inclined.

OpenAI was previously discovered to have made private user conversations with ChatGPT indexable by Google and thus exposed to the world. They quickly patched the flaw when it was discovered, but to say that this is utterly amateur behaviour is an understatement considering the company is supposed to be worth billions of dollars, and employing the smartest engineers on the planet.

OpenAI has also admitted to scanning users’ conversations for “harmful content”, and will report certain conversations to the police if they feel it is warranted. Whether this also applies to the inputs and outputs handled through a third party service such as Genio is unknown, and thus this is highly worrying.

In the past, OpenAI was ordered to preserve user data by a court in New York, despite its policy that such data would be regularly deleted. Although this is not likely to apply to user data in the UK, it begs the question: data passing through its servers can easily be retained by legal means, and as such, its assurance that it will always be deleted is suspect.

Genio have also done their due diligence in maintaining strict agreements with their third party providers that user data cannot be used for AI training purposes. We approve of this, and appreciate their efforts, but we have been in the tech industry too long, dear readers, and we know well the tricks of the trade. We went to OpenAI’s Enterprise Privacy page and found this interesting snippet:

Does OpenAI review my business data?
We may run any business data submitted to OpenAI’s services through automated content classifiers and safety tools, including to better understand how our services are used. The classifications created are metadata about the business data but do not contain any of the business data itself. Business data is only subject to human review as described below on a service-by-service basis.

They may not use it to train a model, but there are many other ways to make use of incoming data, and creating metadata records which will then not be deleted, and likely used for their own business purposes, is something we certainly expected. Believe us when we say this, friends: sometimes the metadata is just as valuable as the data. We may make some educated guesses on what this metadata is and what it may be used for (and of course we hope that its purpose is benign, limited to what is required to maintain the integrity of the system) but, once again, we do not trust Microsoft and OpenAI.

So, let us sum up.

We would venture that no one among the student body was actually asking for this. The use of Genio creates a number of obligations that did not exist before for students and for UBC:

• Students must have a device capable of running the app.
• Students must sign up to the service and hand their personal information to a third party.
• Students must be able and ready to switch it on when needed.
• Students must learn how to use it, or be taught how to use it.
• Students must have an active internet connection in order to use essential features.
• Professors and other students must accept that they are being recorded.
• Professors must accept that their material may be fed into an LLM and distorted by hallucinations, if the student makes use of the AI features of the app.
• UBC must now offer technical support for the app.

In comparison, a human note-taker is just… there. Taking notes. Being adaptable to the student who needs them.

Now, let us be clear, dear readers: everything we have described up to this point is the most generous interpretation of the various actors in this particular play. We would hope that our usual rank cynicism is not, in fact, prophetic of how it will all end. We may believe that Genio, and UBC, have noble intentions. (Microsoft and OpenAI receive no such grace, obviously.) But our instinct is that this entire endeavour will turn into a horrific mess that will inevitably punish students who require these accommodations, making their college life just a little bit worse in the name of progress. Transcription software, though it may be a good use case for machine learning, is notoriously inconsistent with different accents and languages. Genio’s legacy iOS app is reviewed as being glitchy, and its current iOS app has too few ratings to have a score or review at all. (The current app also has no information on accessibility features, which is a damning oversight considering it’s used for accessibility.) The service is far less flexible than a human, and, if its functionality does not work for a given student for whatever reason, they have no recourse except to deal with the issues alone.

Dan McQuillan has written extensively on the use of “disability-washing” of AI, and he has said it far more eloquently than us:

“They want disabled people to have ChatGPT instead of having human rights.”

The belief that AI-assisted tools are the solution to systemic issues faced by marginalized groups is asinine. The students of UBC who require note-taking assistance are not guinea pigs. The conversation would be quite different if Genio were an optional service offered alongside human note-takers, but that is not the case. The choices made by UBC in this instance show clearly that they are not taking the effects of AI seriously, and they do not think their students are worthy of having agency.

We question exactly who made this decision, and who was consulted on it. We wonder how much UBC will spend, in man hours and money, on being locked into Genio’s platform. And we wonder whether the students of UBC will be able to adapt to this indignity. We can only wait, and watch.

We will report on further updates, should any arrive.