iiinova — where your digital self keeps evolving

Cameron Blevins @cblevins.bsky.social · 28/08/2026

We desperately need solutions beyond "online asynch is toast." For years I've told colleagues that AI detection tools are unreliable and problematic. But if Pangram's accuracy results really are this robust, I think schools should think about getting a license: papers.ssrn.com/sol3/papers....

Screenshot from paper by Jabarian and Imas, "Artificial Writing and Automated Detection"

6 Conclusion
We evaluate the current batch of dominant AI-text detectors using a 1,992-passage multi-
genre corpus in conjunction with four advanced LLMs. Three key findings are obtained.
First, a clear ranking in detectors emerges. Pangram maintains near-perfect accuracy
across long and medium length texts. It achieves very low error rates even on shorter
passages and ‘stubs.’ These results are robust to exogenous changes in the threshold and the
use of ‘humanizers’ such as StealthGPT. The other two commercial detectors, Originality.AI
and GPTZero, perform well on long and medium length texts, but, depending on the
detector, struggle on short and ‘humanized’ passages. The open-source RoBERTa baseline
both misclassifies human-written texts and fails to identify up to 51% of AI-generated
passages. Second, recalculating vendor fees as cost-per-true-positive (CPTP) shows that
Pangram is substantially cheaper to implement than both OriginalityAI GPTZero. Third,
the use of policy caps facilitates cross-detector comparisons and allows the policy designer
to select a detector based on a straightforward optimization of their objective function.

2161 0

Cameron Blevins @cblevins.bsky.social · 28/08/2026

Online classes are: a) the modality MOST vulnerable to agentic AI, b) the MOST labor-intensive in terms of faculty-to-student ratios, and c) taught by instructors with the MOST precarious employment. We need to keep this in mind when urging instructors to "teach with intentionality"

Screenshot of MIT report on AI and teaching.
2.5. Teach with intentionality
Most classes taught at MIT will require careful review and, in many cases, substantial adaptation to survive an encounter with Al without serious disruption. Al will undermine the pedagogical assumptions that guided the choice of content, the methods of instruction, and the ways of assessing student performance.
In short, it's time to reflect deeply and extensively on what we teach and how we teach it. Future class structures, assessments, and policies must be built with high intention and a clear sense of purpose, not merely tweaked in reaction to the immediate realities of Al. This does not mean, however, that every course must change in the same way.

34913 0

Cameron Blevins @cblevins.bsky.social · 28/08/2026

Similar result with using ChatGPT to work in Canvas to complete an unessay: it read instructor feedback on an earlier assignment, found primary sources and photos, made a zine-style pamphlet for a campus walking tour, and wrote a process reflection about "my own experience of moving through" campus

First two pages of an campus walking tour pamphlet produced by ChatGPT as part of an unessay assignment.

Screenshot of ChatGPT-authored reflection piece accompanying an unessay project:
Research and Process
I began with my own experience of moving through the Auraria Campus and realizing that its classrooms, lawns, and walkways make the former neighborhood difficult to imagine. My first idea was a general walking-tour map, but the professor's feedback pushed me to choose a specific method and a more analytical message. I decided to make a digital collage-style zine that can be read on screen or printed, folded, and carried while walking. The format combines a simplified route map, five short stops, archival material, original diagram-like illustrations, and questions that ask the reader to look closely at the present landscape.
For research, I first built a timeline from the Denver Public Library neighborhood history and the Auraria Library's displacement and scholarship page. I then used Brian Page and Eric Ross's scholarly article to understand Auraria as part of postwar urban renewal and to see community resistance as a force that shaped preservation and later politics. The 1969 Citizens for Auraria pamphlet became my main primary source because its design and wording show how supporters framed the bond as inexpensive, modern, and beneficial to everyone. I compared that public message with later estimates of the number of people and households displaced. I also used CU Denver's account of the expanded scholarship to connect the 1969 decision to a present-day institutional responsibility.

1134 0

Cameron Blevins @cblevins.bsky.social · 28/08/2026

Agentic AI is also kryptonite for online teaching. Here's Claude's Chrome plugin navigating a dummy Canvas shell in Student Mode to take notes on a lecture and a primary source packet, then use it complete (and ace) a quiz that I tried to ground deeply in that material.

13410 0

Cameron Blevins @cblevins.bsky.social · 28/08/2026

Nearly all of the report's recommendations rest on an assumption that students should learn entirely in-person. This makes sense for MIT, and tbf the authors weren't trying to speak for all of higher ed. But this is simply not how most students receive a college education today.

Screenshot of MIT report on AI and education:
3.2.1. Define and communicate the value of residential
education
As a residential institution, MIT needs to clearly define the value of bringing students to campus and engaging avidly in face-to-face discussions, dialogue, and problem solving. No one reading this report was educated without these basic human interactions. But we cannot take for granted that incoming students will appreciate their value.

1262 0

Cameron Blevins @cblevins.bsky.social · 28/08/2026

This MIT report on AI and teaching is rightfully getting lots of praise: aiandeducation.mit.edu/report/. There's a ton to like, but it exemplifies a trend I see on Bluesky around AI + pedagogy conversations: they tend to leave out online learning entirely.

Screenshot of MIT report on AI and education:
3.1.4. Build structured in-person social learning
into subjects
Because Al tools are disrupting longstanding patterns of social learning at MIT, every subject should include a regular in-person social component (not just sitting in lecture or recitation and quietly taking notes). Instructors should intentionally structure such interactions to achieve desired learning objectives and maintain quality, even in large classes.

69637 0

Cameron Blevins @cblevins.bsky.social · 28/05/2026

All 11 stops on the tour have a corresponding page on our website with multiple ways to listen (embedded audio, links to Spotify, Apple, etc.), written transcript, map of the stop's location, and photo of where to stand while listening: aurariahistory.org/tour/

Screenshot of Stop 2: Growing Up in Auraria, from Listening to Auraria's Past website. It includes navigation buttons, a photo, map, short description, listening options, and an expandable transcript of the audio.

120 0

Cameron Blevins @cblevins.bsky.social · 05/05/2026

Google Slides' new "Beautify this slide" AI tool is comedic gold. Don't think I could have made an uglier, more illegible slide if I tried:

Results of clicking "Beautify this slide" for Google Slides.
Text of slide:
Please fill out your FCQ
for this class by 11:59PM tonight!
https://colorado.campuslabs.com/
Currently only 65% response rate - if you get to 85% everyone gets 4 extra points on Final Exam
Complete assignment: “Did you fill out the FCQ for HIST 1362?” on Canvas (part of Participation grade)

6367 0

Cameron Blevins @cblevins.bsky.social · 16/04/2026

Generative AI has forced me to think a lot more concretely about all the discrete components of doing historical work - and where the value of those different pieces lie

I think we need to stop asking whether Al is "good" or "bad" for historians, or trying to discuss how to police our students who we think are "cheating", and start asking more specific questions. For which of our many tasks has Al already crossed the threshold of
"good enough"? Which parts of our labour process are being cheapened, which are being reorganised? And what happens to the formation of future historians if the junior work through which expertise is built gets outsourced to machines?

1148 0

Cameron Blevins @cblevins.bsky.social · 13/04/2026

Ex. A teacher uses Canvas's "Intelligence Insights" tool to surface "Students in Need of Attention." Technically, it's an AI-powered tracking system. But that's not obvious. Is a teacher expected to a) recognize this, and b) disclose it? community.instructure.com/en/kb/articl...

Example from Canvas feature Intelligence Insights showing dashboard of Students in Need of Attention

031 0

Cameron Blevins @cblevins.bsky.social · 09/04/2026

Interesting parallels here in the “flow” of vibe coding vs. gambling

One thing many of us love about computer programming is our experiences of flow. On the surface, vibe coding can seem to induce a similar flow. However, it often violates the same characteristics of flow that fail with gambling:
• Vibe coding does not provide clear clues of how well one is performing (and even provides misleading losses disguised as wins).
• The match between challenge level and skill level is murky.
• It provides a false sense of control in which people think they are influencing outcomes more than they are.
With vibe coding, people often report not realizing until hours, weeks, or even months later whether the code produced is any good.

230 0

Cameron Blevins @cblevins.bsky.social · 30/03/2026

I did some light iteration for markers, bar charts, etc. but not much on palette. I was struck that it ended up having a similar color scheme as my book cover (although I might be reading too much into it) +

Book cover for Cameron Blevins, Paper Trails: The US Post and the Making of the American West

Screenshot of Cameron Blevins visualization "How Fast Was the Mail"?

100 0

Cameron Blevins @cblevins.bsky.social · 27/03/2026

On transcription: traditional OCR might have transcribed the following as "rmont. lontpeller". Gemini not only recognized this as the Vermont city of "Montepelier" but - crucially - transformed the entire messy table of cities into a usable spreadsheet of data +

Excerpt from 1882 mail transit table showing a difficult-to-read section for the city of Montpelier, Vermont

160 0

Cameron Blevins @cblevins.bsky.social · 27/03/2026

It's a handy tool for researchers (historians, genealogists, stamp collectors, etc.) who are curious how long a letter took to reach its destination. Click or search ~130 cities to quickly see how they fit within a national system of information and how those flows changed over time +

1121 0

Cameron Blevins @cblevins.bsky.social · 18/02/2026

An aside: weekly archival research trips and 2x week in-person classes is what I WISH the job looked like, but the reality for most academic historians is: minimal travel funds, online asynch courses, and lots of "computer work" tasks. This can't be our defense of what makes our jobs AI-proof +

Excerpt from Benjamin Breen, "What is happening to writing?":
"When I think of Al-proof jobs, I think of people like electricians, plumbers, or the surf instructors of Santa Cruz. But I also think about history professors and anyone else whose output includes some combination of in-person engagement and travel-based or otherwise embodied work in a regulated industry. No less than a surf instructor, historians are performing physical services in the real world, although we don't tend to think of it in those terms. We are going into parish church basements to read through baptismal records, finding weird old non-digitized books in rare book shops, piecing together who called Margaret Mead on a certain day in 1954 by reading through her secretary's notes. These are not the everyday tasks of a historian's life, but they are the kind to things we might do, say, once a week. Couple that with twice a week in-person classroom time, and I simply flatly disagree with anyone who thinks this combo will be replaced by a Sonnet 4.6-type model, no matter how good it gets at creating Excel spreadsheets, translating Latin, or explaining linear algebra."

131 0

Cameron Blevins @cblevins.bsky.social · 13/02/2026

So much for soup to nuts. Next I gave it a transcribed version of the data I had typed into a CSV file years ago. 7 minutes later, I had a fully interactive web map. I hit my rate limit before I could make any further tweaks but even this first autonomous pass is pretty stunning: +

140 0

Cameron Blevins @cblevins.bsky.social · 13/02/2026

Womp womp. Complete failure. It kept trying to use OCR packages to extract the data rather than vision + reasoning, which was never going to work with this kind of messy tabular source. I finally shut it down after 30 min. of spinning its wheels with gibberish OCR: +

Poor OCR'd version of mail transit table from 1882.

130 0

Cameron Blevins @cblevins.bsky.social · 13/02/2026

"Just point Claude Code at some files and see what it can do." I wanted to test this on a historian's workflow so I used a small 4 pg. PDF from my dissertation research that never made it into my book: an 1882 table of how long it took the mail to travel between different cities. +

Sample page from a table from 1882 postal guide of time of transit of mails between major US cities

251 0

Cameron Blevins @cblevins.bsky.social · 13/01/2026

I then used Gemini 3.0 to transcribe that same handwritten section. Gemini made exactly 2 errors (WER: 2.70%, WIP: 96.47%), and one of those "errors" was actually a word that Mujic seems to have misidentified. Translation: a nearly flawless transcription. +

100 0

Cameron Blevins @cblevins.bsky.social · 13/01/2026

I tried out @dougboyd.bsky.social's transcription comparison tool to quantify this early-ChatGPT transcription vs. Mujic's transcription. Nearly 1/2 the words needed correction, and only 30% of the content came through reliably (WER: 46.85%, WIP: 29.31%). Translation: hot garbage. +

Screenshot of transcription comparison metrics.
WIP (Word Information Preserved): 29.31% — higher is better
WER (Word Error Rate): 46.85% — lower is better
Counts: Reference N=111 • Machine M=107 • Matches C=59 • S=48 • D=4 • I=0
Precision/Recall/F1: 100.00% / 93.65% / 96.72%
Reading the numbers:

WIP rewards overlap and penalizes missing/extra wording. 100% = perfect match.
WER is standard in ASR: (S + D + I) ÷ N. Lower is better.
S/D/I reveal whether misses come from drops, extras, or swaps.

120 0

Cameron Blevins @cblevins.bsky.social · 13/01/2026

Mujic provided a super helpful deep dive on one particularly messy section of the poem riddled with insertions and crossed-out words, then included a side-by-side comparison of their transcription vs. early-ChatGPT. I decided to take a closer look to try and measure the errors in this section. +

Section of a handwritten 1864 poem, from Julie Mujic, "Generative AI and Civil War Primary Sources: A Methodology Experiment"

100 0

Cameron Blevins @cblevins.bsky.social · 17/10/2025

Kick off your morning at #WHA2025 with some spicy takes on Generative AI and history from Sean Fraga, Rachel Birch, @jasonheppler.org, @regan008.bsky.social, and myself (8:15-9:45am, Santo Domingo Room)

Roundtable:
Western History in the Generative AI Era
Sponsored by the WHA Digital Scholarship Committee
Location: Santo Domingo (Convention Center)
Chair: Sean Fraga, University of Southern California
Cameron Blevins, University of Colorado Denver
The State of Generative AI in October 2025
Rachel Birch, George Mason University
Mowing the Lawn with a Hammer: Pivoting from
Codependency to Cointelligence with AI Tools
Jason Heppler, George Mason University
Tools with Thought: Generative AI as an
Assistant in Historical Research
Amanda Regan, Clemson University
Bridging the Gap: Leveraging AI Tools to Lower
Barriers for Historians Learning to Code
Comment: Audience

071 0

Cameron Blevins @cblevins.bsky.social · 19/07/2025

What a memory @scottbot.bsky.social! The post office department occasionally published transit time for mail between major cities, that’s the closest I can think of. Ex. This one from 1882. Happy to dig up the other ones I have if you want to email me!

110 0

Cameron Blevins @cblevins.bsky.social · 16/07/2025

A lot of LGBTQ scholarship has focused on: a) gay men, and b) big cities. By mapping 5,000+ locations across 26 issues from 1975-1989, we showed something different: a much more expansive and dispersed geography of lesbian readers that tilted towards rural areas and mid-sized cities

Map of locations printed in a subset of issues from Lesbian Connection between 1974 and 1989 in the contiguous United States, sized according to the number of times they appear.

120 0

Cameron Blevins @cblevins.bsky.social · 26/12/2024

Can’t imagine what you’re talking about here

160 0

Cameron Blevins @cblevins.bsky.social · 02/12/2024

Taking a break from grading on #DayofDH2024 to redesign our Data Advocacy For All website. I know zero Javascript/CSS, but used a combo of Claude + Github Copilot to get a fully functioning, clickable button interface for filtering and searching resources.

Screenshot of Data Advocacy For All website section using a filter and search interface, based on Resource Type, Literacy Domain, and Literacy Subdomain.

070 0

Cameron Blevins @cblevins.bsky.social · 20/11/2024

The state-level mortality figures are worth a look, although I'd caution that many individual states have very, very wide confidence intervals. From their full paper: www.pnas.org/doi/10.1073/...

Excess mortality chart for states during the US Civil War (https://www.pnas.org/doi/10.1073/pnas.2414919121)

021 0

Cameron Blevins @cblevins.bsky.social · 29/10/2024

New post, “A Large Language Model Walks Into an Archive…” explores how historians can use off-the-shelf LLMs to work with primary sources 📜 cblevins.github.io/posts/llm-pr...

Page one of a handwritten letter by Benjamin Curtis (1886)

Transcription of a handwritten letter by Benjamin Curtis (1886) completed by ChatGPT

48331 0