Marc Watkins

YouTube video by Josh Brake Comet Goes to School

AI is now browser based. You can view a video of Perplexity's Comet completing an online assignment from Josh Brake. Don't panic. The speed is the giveaway. It is good practice to start checking LMS timestamps for how long students access and complete assignments youtu.be/opOx3Msw5tI?...

Beyond the many moral and ethical objections I raised about using AI to grade, here's a concrete one: AI grading will never be secure because of prompt injection. Simply hiding a phrase within an assignment is all it takes to bypass even the most advanced AI models.

The Dangers of using AI to Grade Nobody Learns, Nobody Gains

marcwatkins.substack.com/p/the-danger...

Some teachers are using AI to grade and few are talking about it. Scantrons are one thing, but automating grading, feedback, and assessing written work with AI poses massive ethical challenges and should not become the default or go to choice in education.

👇 in comments

Bluesky is blocked in my home state. I can only access it via VPN or when I’m traveling as I am now.

More college students are using AI for class. Their professors aren't far behind More college students are using AI chatbots to help them with their studies. But data recently released by an AI company shows they're aren't the only ones using the technology.

I was on NPR Morning Edition to talk about the deeply troubling ethical issue of faculty using AI to grade student assignments. I'm glad this issue is getting more attention, but it is also one I don't see clear a path forward. www.npr.org/2025/10/07/n...

The Hallucinating Machines We Can't Live Without An Algorithm Brought You To This Anti-Algorithm Essay

Our relationship with new technologies has always been fraught and AI isn’t new in that regard. After 3 years of ChatGPT, over a decade with social media, and screens in our pockets, are we able to chart a path forward or helplessly lost in hype and doom?

marcwatkins.substack.com/p/the-halluc...

Age Verification is Coming for Chatbots I’m now more convinced than at any previous point that governments will start regulating generative AI via limiting access to tools like ChatGPT based on a user’s age, and I’m not sure that’s a good t...

Popping in via VPN to share how age verification is actively limiting access to services like Bluesky. I think AV will eventually impact access to AI. I don't think AV effectivly protects children and don't believe this is the future any of us wants. marcwatkins.substack.com/p/age-verifi...

It’s been real, but this is likely the end for me on this platform. I have to use a VPN to post or read contact because of the state I live in. This is chilling and should be a wake up call for what’s coming.

I guess I won’t be posting here much longer. Being in a state that’s instituted age verification laws causing social media companies like Bluesky to restrict access isn’t an isolated event—it’s where the internet is headed. I’ll use a VPN for now, but that’s not going to be sustainable.

I think it's valuable, but the real skill set that matters remains separating the bullshit from what actually works. Social media has created bizarre frameworks for how people gain knowledge.

Higher Education Needs Frameworks for How Faculty Use AI The anticipation surrounding GPT-5 created a dizzying buzz among tech enthusiasts who were obsessed with the idea that true artificial intelligence, known as AGI, was imminent.

marcwatkins.substack.com/p/higher-edu...

Faculty guidelines and frameworks for using generative AI with their students are often unclear and vague. The institutional silence around faculty AI usage risks sending a message of indifference instead of charting a clear path forward. I created a VALUES framework for faculty to consider.

Locking this up and potentially taking down my post because 1) I am an AI skeptic 2) people aren’t interacting with the post in good faith. It’s always right to call out AI hype. The point was AI isn’t going away if the hype dies and being overly dismissive isn’t helping us navigate it.

Locking this up and potentially taking down my post because 1) I am an AI skeptic 2) people aren’t interacting with the content in good faith. It’s always right to call out AI hype. The point was AI isn’t going away and being overly dismissive isn’t helping us navigate it.

I hope people dunking on OpenAI for how crappy GPT-5 is and how hyped AGI has been take a step back and consider how the LLMs we have today will take society years to fully absorb. This technology is impact our world in ways we can barely comprehend. AI isn't going away even if the hype does.

A little over a year from now, the first generation of ChatGPT students will graduate college with no guidance about how to use or resist AI in their careers. A college freshmen will arrive on campus having 4 years of access to a chatbot and we think bluebooks are an answer?

The AI Wolf That Education Must Face GPT-5 is now here and freely available to all.

open.substack.com/pub/marcwatk...

Another semester approaches and faculty face a super charged version of ChatGPT. I want to talk about value and meaning with my students, not necessarily AI. Here are some ideas to do just that. I also have a session submitted to SXSW Edu about AI aware teaching that I’d love for you to vote for. 👇

Will Microsoft update the education version of Copilot with GPT-5? This was formerly known as Bing Chat Enterprise and hasn't been updated with a new model since GPT-4 launched in 2023. That would give data protection w/ GPT-5 to a lot of schools.

I stand corrected. They simply moved it to the "+" feature. I missed it under the . . . more section

RIP ChatGPT's Study Mode. It lasted ten days as a feature before GPT-5's launch removed it. Now students will need to ask the model to not complete tasks for them manually. The point was a ⏹️ button students could use to create a socratic tutor, a space where the model wouldn't offload learning.

But when you attach a file and ask it very specific questions, it does a fairly decent job of grounding its responses in the data. Here's a summary of the Walton Foundation report on AI saving teachers six hours a week. It took several minutes to run this so I'm not sure how practical that is.

4/4

Without search, the model just makes up material. Wildly so. Asking for a list of publications from me and none of these are real. We're back to the good old days of early GPT-3 and ChatGPT.

3/4

I cannot overstate how bad hallucinations are without RAG. Asking the model for the current AI outlook on my campus produced no factual information, which isn't surprising given the model's size.

2/4

Giving oos 20b a second look with LM Studio, which is much better at running it. You can toggle the reasoning settings and this does improve the output. Medium setting seems to be the sweet spot, but on high you have to raise the context limit to get the best output to count Rs in strawberry.

1/4

3️⃣ Fail: I asked it to code a game on climate change. The model began searching the web then got stuck in an endless cycle of web search and reasoning. We never got to tools to code or refusal, just a loop in search. After 20+ minutes I ended the chat without a result with a much lower battery.

3/3

2️⃣ Partial success: I used search and asked about the current situation in Gaza. Multiple errors. The model took 8 minutes to come up with a response and gave me a single sentence response about death toll, not a summary of the overall situation:

2/3

I have a Macbook Pro/ M4 48 gbs and ran GPT-oss 20b via ollama and these were the results on three tests for reasoning and search. My overall impression is the usefulness of the model is very limited on a personal device.

1️⃣ Fail: the strawberry test

1/3

Read 4.5 in the model card carefully— OAI did not evaluate the models search function related to hallucinations. That oss can use tools doesn’t mean it is effective in using search to ground responses. Kimi K2 is a great example of a powerful open model that both uses tools and hallucinates wildly.