When ‘AI’ Becomes an Accusation: Accessibility, Authorship and the Problem With Treating AI-Assisted as AI-Generated

The world is changing at a rapid rate, and AI seems to have become another centre of division across industries and communities.

AI can certainly feel new and frightening. We are seeing so many changes, concerns and genuinely negative consequences unfolding within a very short window of time. It is debated on the news, it is increasingly part of the news, and social media platforms are now openly incorporating AI into their products and policies.

Artificial intelligence is everywhere.

It is in our social media platforms, our design software, our phones, our search engines, our writing tools and, increasingly, in accessibility software that many people rely upon to communicate, study and work.

Yet our conversations about AI often remain remarkably binary.

Something is either “human-made” or “AI”.

A person either “wrote it themselves” or they “used AI”.

And increasingly, simply suggesting that AI may have been involved in somebody’s work can carry an implication that they did not really create it at all.

I think we need to have a much more nuanced conversation.

It may seem strange for a dog behaviourist to be writing about this — and, in typical me fashion, carrying out far more research into it than I probably originally intended — but this has become personal.

Instagram has recently been labelling not just my work, but the work of many other creators, as involving AI.

Those who have known me for years know that I have always had a love/hate relationship with social media. Facebook is somewhere I largely exist through my business profile to protect my mental health. Instagram has irritated me for years, long before generative AI entered the conversation. First there were arguments over video and music rights when the music had legitimately come from another platform. Then came more and more features for my neurodivergent brain to navigate.

And now, it is labelling my posts as AI.

Firstly, that label does not accurately describe how my work is created. Secondly, it is not happening only to me. Other creators have reported similar problems, to the point that people are even sharing workarounds such as taking a screenshot of an original image before uploading it in an attempt to avoid automated labelling.

At the same time, Meta is actively integrating more AI tools into its own platforms, while Substack has introduced its own AI-detection system and has also labelled some of my writing as potentially AI-assisted.

Being autistic and having learning difficulties changes how I experience all of this.

Last year, I had a couple of incidents on Reddit where somebody referred to me as “ChatGPT”.

For somebody who has spent a lifetime being given labels by other people, being reduced to “ChatGPT” because of the way I write and communicate was extremely painful. The same discomfort now sits behind my experiences with Instagram and Substack.

So I no longer post on Reddit, and I have found myself questioning whether I want to continue posting on Instagram either.

What made one of those Reddit comments particularly frustrating was that the post itself explicitly explained that the physical materials for my classes would be sent through snail mail.

The photograph showed my shelf. My mushroom agate crystal. Physical products that I had spent many, many hours creating, printing, assembling and building, photographed on my own shelving unit.

I had edited the photograph using Adobe, as I have done with my images for years, and the branding that appears across my work is not stock artwork or something generated from a prompt either. I draw my own branding and logo in Procreate.

These are ordinary parts of my creative process — photographing, drawing, editing, designing and assembling — yet the finished result was still dismissed as AI.

So, in typical me style, I started researching why this is happening.

This is not an argument that all uses of artificial intelligence are ethical, harmless or desirable. They are not.

Generative AI raises legitimate concerns around misinformation, environmental cost — especially as an advocate for animals and biodiversity — intellectual property, bias, automation, data provenance, manipulation and the increasing volume of low-quality automated material online (Chien et al., 2026).

Researchers are already examining what Boatwright, DiRusso and Pyle (2026) describe as “AI slop”: the rapidly growing volume of AI-generated content and accounts that can contaminate social-media datasets and make it increasingly difficult for researchers to determine whether they are studying human communication at all. Their concern is important and should not be dismissed.

But recognising those problems does not require us to treat every interaction with AI as equivalent.

There is a substantial difference between generating hundreds of automated posts, asking a machine to manufacture expertise that somebody does not possess, using AI to impersonate another person, asking a system to write an article or book that the supposed author has never researched or critically considered — and a human being using technology to help them express thoughts, knowledge and ideas that remain their own.

That distinction matters.

And this distinction is not limited to writers, creators or students.

AI is already being incorporated into active scientific and mathematical research. A September 2026 New Scientist special report described mathematicians and researchers using AI agents to explore extraordinarily difficult mathematical problems, including work surrounding the Navier–Stokes equations. AI systems were able to produce possible solutions, counterexamples and avenues for investigation at extraordinary speed, while human mathematicians were still needed to examine, formalise, interpret and understand what those systems had produced (Sparkes & Padavic-Callaghan, 2026).

Perhaps that is another useful reminder that using AI does not automatically mean surrendering human expertise.

In research, as elsewhere, the more meaningful questions are not simply whether AI was involved, but what role it played, what the human contributed, and how its output was scrutinised.

Many professionals are also now using AI-enabled tools as an ordinary part of remote work. Zoom's AI Meeting Summary uses speech-to-text data to generate meeting summaries, while tools such as Fathom can record and transcribe meetings before creating AI-generated summaries from those transcripts. Zoom separately describes live captions as a real-time accessibility feature (Zoom, 2026a, 2026b; Fathom, 2026).

I keep captions switched on during my own Zoom meetings because sometimes my brain hears something differently from what was actually said. Being able to read the words as well as hear them helps me process the conversation accurately.

The technology assists my access to the conversation.

It does not have the conversation for me.

AI-enabled does not mean AI-generated

One of the problems with our current language is that AI is no longer a separate programme that somebody consciously opens when they decide to “use AI”.

AI functionality is increasingly embedded inside ordinary software.

Canva is one example.

Canva now contains a dedicated AI assistant that can generate and refine designs, brainstorm ideas, work with text, images and voice, and help users polish existing material. Canva describes the user as remaining in creative control, and the AI assistant must itself be accessed within the software (Canva, 2026a).

Importantly, Canva's own AI Product Terms distinguish between material supplied by the person — the Input — and material generated by its AI Products — the Output. Its rules concerning AI-generated material apply specifically to the use of those AI Products (Canva, 2026b).

Simply opening Canva therefore does not mean that Canva created your work, and an Instagram AI label cannot, by itself, tell us which Canva functions were actually used.

A person can upload their own photographs, create their own illustrations, choose their own typography, arrange every element manually and write every word themselves while working inside software that also happens to contain AI features.

Adobe presents a similar picture.

Photoshop now contains specific generative-AI tools alongside traditional editing tools. Adobe's own guidance distinguishes generative functions such as Generative Fill from the wider Photoshop environment and states that Content Credentials may identify whether material was generated or modified using generative AI (Adobe, 2026a; Adobe, 2026b).

Wix provides another example from my own working life. I use Wix to build and maintain my website, and as AI functionality became increasingly integrated into the platform I found it beginning to generate draft blog content for me that I had not individually asked it to write. Wix's current AI Marketing Agent is designed to proactively create marketing plans and content, including draft blog posts, although Wix states that generated content must still be reviewed and approved before publication (Wix, 2026a).

I turned those AI functions off.

That decision was not because I believe that the involvement of AI automatically makes something unethical. It was because the generated writing was not my voice, and some of what it produced was inaccurate. For me, that crossed an important line between technology assisting me to communicate what I already wanted to say and technology deciding what I should say in the first place.

Perhaps that is another reason why simply asking whether somebody “uses AI” tells us so little. I use some forms of technological and AI assistance because they remove barriers or help me organise and communicate my own thoughts. I deliberately reject others when they begin replacing the part of the process that I want — and need — to remain mine.

An AI-enabled piece of software is therefore not synonymous with AI-generated work.

This distinction will only become more important.

Meta itself is now actively embedding AI into the tools used by creators and businesses. Its September 2026 launch of Meta One describes increased access to AI-powered image and video creation, editing, Restyle tools and other creative functions across Instagram, Facebook, WhatsApp and Meta AI (Meta, 2026a).

AI and machine-learning functions are appearing across technologies many of us already use, from voice assistants and search engines to photography, communication and writing software.

But even here, precision matters.

Automation itself is not necessarily AI.

A self-checkout, for example, may be automated without generative AI having anything whatsoever to do with the transaction. If we use “automated”, “algorithmic”, “machine learning” and “generative AI” interchangeably, we recreate exactly the same problem this article is trying to challenge.

It becomes increasingly difficult to defend a binary division between “people who use AI” and “people who do not” when AI functions are being integrated into everyday creative and professional workflows.

A more meaningful question is:

What did the technology actually do?

Accessibility complicates this conversation further

This issue becomes especially important when technology is used to remove barriers.

Read&Write is an established literacy-support tool. Its Disabled Students' Allowance resources describe tools including text-to-speech, screen masking, spelling support and other functions designed to help students understand information and express themselves (Texthelp, 2026a).

Its workplace guidance describes Read&Write as particularly useful for people who are neurodivergent and for people whose conditions affect reading, writing, working memory, fatigue or dexterity. Among the features it identifies is dictation (Texthelp, 2026b).

Talk&Type allows someone to dictate their written content.

Think about what that means for authorship.

A person — me included — could speak an entire article aloud while software transcribes every word onto the screen.

Technically, the computer has entered every letter.

But did the computer author the article?

Of course not.

The ideas, language, decisions and meaning originated with the person speaking.

Read&Write now also contains newer AI functions. Its Simplify AI feature explicitly uses OpenAI to rewrite material using simpler words and sentence structures (Texthelp, 2026c).

Within a single accessibility programme we can therefore find traditional assistive technologies, predictive tools, speech recognition and generative AI.

Calling the entire programme simply “AI” tells us almost nothing about what somebody actually did with it.

And for people who use accessibility technology, this distinction is not merely academic.

Technology can provide the bridge between having something to say and being able to communicate it effectively.

Assistance with spelling does not remove authorship.

Speech-to-text does not remove authorship.

Using text-to-speech does not mean somebody has not read or understood the material.

Help reorganising a difficult sentence does not automatically mean that the ideas within that sentence belong to somebody — or something — else.

Grammarly is another interesting example.

Millions of students use Grammarly, and its current education tools now include AI-assisted writing features alongside spelling, grammar and clarity support. Grammarly has even developed an Authorship feature that attempts to make the process of writing visible by distinguishing between material typed by the person, pasted material and AI-assisted or AI-generated text (Grammarly, 2026a; Grammarly, 2026b).

That is an important development.

Even a company providing AI writing tools is recognising that how somebody produced a piece of writing carries information that the finished text alone cannot reveal.

We need to be very careful that anxiety about generative AI does not turn into suspicion of the very technologies people use to participate.

What does the research say about AI assistance?

A 2026 experimental study by Møller, Romero, Jurgens and Aiello examined generative-AI tools in a simulated social-media environment involving 680 US participants.

Participants were assigned either to a control group with no AI or to conditions providing different forms of assistance, including chat, conversation starters, feedback on drafts and suggested replies.

The researchers found a complicated picture rather than a simple good-versus-bad result.

Some forms of AI support increased participation and the amount of content people contributed. At the same time, AI-assisted discussions could be perceived as less authentic or lower quality, particularly where contributions felt generic or impersonal (Møller et al., 2026).

That matters.

I do not want a world filled with generic machine-produced communication either.

But another finding is particularly relevant to the way we talk about authorship.

The researchers explicitly distinguished between material directly copied from generative AI and more nuanced uses such as revising somebody's own draft or taking inspiration from an AI suggestion. They suggested that directly copied generated content should be disclosed, while recognising that those more nuanced forms of assistance may not necessarily require the same labelling approach (Møller et al., 2026).

In other words, researchers studying AI-mediated communication themselves recognise that:

AI-generated and AI-assisted are not automatically the same thing.

Perhaps even more strikingly, participants were not particularly reliable at judging who had used AI.

In the experimental control group, where participants had no AI assistance, other participants nevertheless believed that 13.8% of them had used AI.

Suspicion rose in the AI conditions, but the false suspicion within the control group demonstrates something important: believing that writing “sounds like AI” is not proof that AI produced it (Møller et al., 2026).

That finding should give us pause before publicly challenging somebody else's integrity based on style alone — especially if we haven't first stopped to ask: Have I actually read the caption, or am I going off an assumption prompted by a social-media bot?

Detection is not authorship

This becomes particularly relevant when social platforms themselves begin labelling content.

Meta has already acknowledged problems with its own approach.

In 2024, Meta explained that its original “Made with AI” labels did not always align with people's expectations. Some images received the label because of relatively minor AI-enabled modifications, including retouching. Meta subsequently changed the wording to “AI info” in an attempt to provide greater context (Meta, 2024).

The issue has not disappeared.

A September 2026 investigation by The Verge documented users reporting “AI Content” labels after relatively minor edits, including use of Canva's Background Remover, while some genuinely AI-generated test images reportedly escaped the label.

This is journalism rather than a controlled scientific assessment, so it should not be treated as an established error rate. It does, however, illustrate why an Instagram label cannot necessarily tell us how much AI was involved or what role it played in creating the finished work (Weatherbed, 2026).

Substack takes a different approach.

Its “Scan for AI text” function uses Pangram to estimate how much text appears human-written or AI-assisted.

That word — estimate — is important.

Substack also allows writers to provide a “How I make this” statement explaining their creative process, disable AI detection on individual pieces, and report detection errors (Substack, 2026).

None of this means detection technology has no value.

It means the result needs to be interpreted for what it actually is.

A detection system may identify patterns associated with AI-assisted text.

Metadata may show that a generative or AI-enabled function interacted with part of a file.

A Content Credential may record that a generative tool modified part of an image.

None of those facts, alone, answer:

Who developed the argument?

Who did the research?

Who supplied the lived experience?

Who decided what was relevant?

Who rejected suggestions?

Who checked the references?

Who corrected mistakes?

Who made the final editorial decisions?

Who meant what the finished piece actually says?

Those are questions of authorship, agency and responsibility.

There are genuine reasons to remain cautious

Nuance must work both ways.

AI can assist communication, but it can also distort it.

Tsirtsis, Rawal, Russell, Mittelstadt and Wachter (2026) examined what happened when large language models were asked to edit human-written material while preserving its meaning.

They found that models could introduce directional shifts when editing writing about contested subjects. Their subsequent mathematical modelling suggested that small biases introduced repeatedly through AI-mediated communication could potentially accumulate across social networks.

This work is currently a preprint rather than a peer-reviewed journal publication, so that limitation matters.

But the principle is worth taking seriously.

If I ask an AI system to help me polish something, I still have a responsibility to read the result.

Does it still say what I meant?

Has a qualification disappeared?

Has certainty been added where I deliberately expressed uncertainty?

Has my terminology changed?

Has a reference been misunderstood?

Has my voice disappeared?

Human oversight should not be ceremonial.

It is part of responsible AI use.

Likewise, Gündüzyeli's (2025) review of 67 peer-reviewed publications found potential benefits of AI and social media for communication, personalisation and organisational resilience while also identifying significant concerns around misinformation, privacy, algorithmic bias and trust.

These technologies deserve scrutiny.

But scrutiny is different from accusation.

Authorship is not measured by how many keys somebody pressed

Perhaps this is the question I keep returning to:

What do we actually mean when we say somebody “created” something?

If I research a subject, read the papers, take notes, form an argument, decide what I believe the evidence supports, write a draft and then use technology to help organise or polish that writing, where exactly does authorship supposedly disappear?

After spellcheck?

After grammar correction?

After dictation?

After word prediction?

After asking software for a synonym?

After restructuring a paragraph?

After asking an AI system whether a sentence communicates what I intended?

There obviously is a point at which assistance can become substitution.

If somebody asks generative AI to research, think, argue and write on their behalf and then publishes that material without meaningful involvement, scrutiny or understanding, that raises very different questions.

But that is precisely why “AI was involved” is an inadequate description.

We need language capable of distinguishing:

human-created work;

AI-assisted human work;

AI-edited work;

AI-generated work;

and automated or mass-generated content.

Those categories can overlap, and technology will continue to blur their boundaries.

Nevertheless, they describe meaningfully different creative processes.

Collapsing all of them into “AI content” loses information rather than providing it.

Accessibility should not have to prove its innocence

There is another aspect of this conversation that concerns me particularly as a student and as somebody with learning difficulties who has used assistive technology for many years — from Dragon dictation software in the past to Read&Write and Talk&Type now.

This is not a fringe issue.

In March 2026, the Department for Education in England opened a consultation specifically examining assistive software funded through Disabled Students’ Allowance and the increasing integration of AI into those tools.

The consultation explicitly recognises that AI-enabled software can be powerful in supporting disabled students while also raising legitimate questions about university AI policies, ethics and software that can prepare academic content on a student's behalf. It asks not only whether DSA should fund software containing generative AI, but also: “In what ways can AI be used to improve support for disabled students?” It then separately asks how general-purpose AI differs from AI integrated into assistive software and how those differences affect accessibility and user experience (Department for Education, 2026).

That distinction matters enormously.

Disabled Students’ Allowance exists to help with additional study-related costs arising from disability and can fund support including specialist equipment, software and non-medical help.

The same government consultation notes that functions once confined to specialist products are increasingly becoming built into mainstream technology (Department for Education, 2026).

The fact that AI functions are now becoming embedded within some of those technologies does not suddenly erase their accessibility purpose.

But it does mean that universities, governments, technology companies and disabled students are having to think much more carefully about what different AI functions actually do.

That is exactly the nuance I feel is missing when somebody sees the word “AI” and immediately treats every use of technology as equivalent.

Long-time readers may also remember that I used my paid NaturalReader account to create accessible versions of some of my academically researched blogs. I would have the written content read aloud while screen-recording it so that people who benefited from listening rather than reading could access the material without needing to purchase the software themselves.

NaturalReader itself describes its text-to-speech tools in terms of accessibility, including features intended to support different reading needs. It too has continued to evolve and now contains AI-enabled features alongside its reading and accessibility functions (NaturalReader, 2026a; NaturalReader, 2026b).

Again, the presence of AI somewhere inside the software does not tell us what a particular person used the software to do.

When we decide that polished language, particular sentence structures or certain creative styles are inherently suspicious, the people most likely to be asked to explain themselves may include those who already use technological support to communicate.

People use assistive technology for countless reasons.

They may dictate rather than type.

They may use software to help with spelling, visual processing, reading or sentence construction.

They may need support organising information.

They may have excellent subject knowledge while finding written expression substantially harder.

They may use technology because fatigue, pain or physical access changes from one day to another.

The quality of someone's ideas should not be measured by how independently their fingers produced the final characters on a screen.

We should absolutely care about authenticity.

But authenticity does not require technological purity.

Transparency needs nuance too

I support transparency around genuinely generated material.

I also think people should remain responsible for what they publish regardless of which tools helped them create it.

If AI invents a citation and I publish it, I am responsible for failing to check it.

If AI changes the meaning of my argument and I approve the revision, I am responsible for what I published.

If I present AI-generated expertise as knowledge I personally possess, that raises obvious ethical concerns.

But transparency should inform people, not erase meaningful distinctions.

“AI-generated” should mean something.

So should “AI-assisted”.

As AI becomes embedded into Meta, Canva, Adobe, accessibility software and everyday communication tools, demanding that we classify a person's entire piece of work according to whether any AI technology touched it at any stage becomes increasingly unhelpful.

What matters is what the technology contributed and what the human contributed.

Who originated the ideas?

Who holds the expertise?

Who made the choices?

Who remains accountable?

And who ultimately owns the meaning?

Perhaps we need curiosity before accusation

There are important conversations to have about artificial intelligence.

We should have them.

We should challenge poor practice, undisclosed generated material, fabricated expertise, misinformation, automation masquerading as genuine engagement and the increasing pollution of online spaces with meaningless generated content.

But we should also resist turning those legitimate concerns into a culture of suspicion.

Seeing something that “looks AI” is not evidence.

Knowing that somebody uses Canva is not evidence.

An Instagram label does not tell us somebody's creative process.

A detection percentage does not tell us who had the idea.

And using technological assistance does not automatically mean surrendering authorship.

Before publicly questioning another person's integrity, perhaps the more ethical approach is also the simplest:

ask.

Because responsible use of technology deserves scrutiny.

But people deserve curiosity before judgement.

And accessibility should never become collateral damage in our attempt to decide what counts as authentically human.

References

Adobe. (2026a). Adobe Generative AI User Guidelines. Updated 15 May 2026.

Adobe. (2026b). Content Credentials overview. Adobe Creative Cloud.

Boatwright, B. C., DiRusso, C., & Pyle, A. (2026). Somebody's poisoned the water hole!

Advancing social media data collection principles to account for the effects of AI slop on public relations scholarship. Public Relations Review, 52(3), 102716. https://doi.org/10.1016/j.pubrev.2026.102716

Canva. (2026a). Canva AI: Your all-in-one AI assistant.

Canva. (2026b). AI Product Terms. Effective 26 June 2026.

Chien, A. A., Gupta, U., Ren, S., Sriraman, A., & Tomlinson, B. (2026). Strategies and design for increasing AI sustainability. Nature Reviews Clean Technology, 2, 624–638. https://doi.org/10.1038/s44359-026-00195-w

Department for Education. (2026). Assistive software funded through Disabled Students’ Allowance: Consultation. Published 26 March 2026.

Fathom. (2026). Fathom Help Centre.

Grammarly. (2026a). Grammarly Authorship: Ensure authentic writing in the age of AI.

Grammarly. (2026b). Free AI writing assistance for students.

Gündüzyeli, B. (2025). The role of social media and artificial intelligence (AI) in enhancing digital marketing resilience during crises. Sustainability, 17(7), 3134. https://doi.org/10.3390/su17073134

Meta. (2024). Our approach to labeling AI-generated content and manipulated media. Meta Newsroom.

Meta. (2026a). Introducing Meta One: A subscription service with more features and AI to create, connect, and stand out. Meta Newsroom, 15 September 2026.

Møller, A. G., Romero, D. M., Jurgens, D., & Aiello, L. M. (2026). The impact of generative AI on social media: An experimental study. Scientific Reports, 16, 9376. https://doi.org/10.1038/s41598-026-40110-8

NaturalReader. (2026a). Free text to speech online for accessibility.

NaturalReader. (2026b). NaturalReader AI Text to Speech: Features.

Sparkes, M., & Padavic-Callaghan, K. (2026). Biggest breakthrough in decades. New Scientist, Issue 3613, 19 September 2026, pp. 5–6.

Substack. (2026). How can I detect AI on Substack? Updated 9 August 2026.

Texthelp. (2026a). Introduction to Read&Write in the DSA.

Texthelp. (2026b). A quick introduction to Read&Write.

Texthelp. (2026c). How to simplify some text using AI.

Tsirtsis, S., Rawal, K., Russell, C., Mittelstadt, B., & Wachter, S. (2026). AI-mediated communication can steer collective opinion. arXiv preprint arXiv:2605.16245.

Weatherbed, J. (2026). Instagram's AI detection is a mess (again). The Verge, 4 September 2026.

Wix. (2026a). AI Marketing Agent: An overview. Wix Help Center.

Zoom Video Communications, Inc. (2026a). Using Meeting Summary with AI.

Zoom Video Communications, Inc. (2026b). Frequently asked questions about meeting captions and transcript updates (May 2026).