iiinova — where your digital self keeps evolving

Andres Karjus @andreskarjus.bsky.social · 13/10/2026

See on see mida AI tihti ütleb kui talle tekst sisse sööta

Reposted by Andres Karjus

Digital Humanities Potsdam @dhpotsdam.bsky.social · 12/10/2026

🤖 The #CodeAndCulture talk series restarts with a talk by @andreskarjus.bsky.social on "LLMs, Big Data and the Agentic Future". Register here: plu.sh/codeandculture

031 3

Andres Karjus @andreskarjus.bsky.social · 12/10/2026

In it, I discuss, among other things... - how #LLM technologies are changing the #DH #CHR #CUDAN research landscape - some best practices - but it's not just data annotation - how will agentic use affect identities of ppl with "digital" or "analytics" job titles - impactful vs incremental science

000 0

Andres Karjus @andreskarjus.bsky.social · 12/10/2026

New preprint out, of a chapter in the upcoming Handbook of Cultural Analytics: "Large Language Models, Machine-Assisted Cultural Analytics, and the Agentic Turn" Feedback and critique very much welcome. Tell me if I missed something important!

120 2

Andres Karjus @andreskarjus.bsky.social · 11/10/2026

That's also the state of journal review/academic publishing. Will probably only escalate until gets so insane that a meaninful change is unavoidable. Which might not be a bad thing in the end.

020 2

Andres Karjus @andreskarjus.bsky.social · 08/10/2026

It is only Màth if it's produced in the Mathagne region, by the Mathematicians' Guild following the Méthode Traditionnelle

040 4

Andres Karjus @andreskarjus.bsky.social · 29/09/2026

These thinking/reporting traces really are nice to watch scroll by (an extremely nerdy kind of joy I know)

000 0

Andres Karjus @andreskarjus.bsky.social · 22/09/2026

So all these nice corpora gated behind bad search interfaces, useless KWIC views, OCR behind scans, browser-only, no download. Now you can just get codex/ccode to dissect the urls/search/display interface until it figures out how to scrape anyway and have it download full datasets? How convenient.

020 2

Andres Karjus @andreskarjus.bsky.social · 09/08/2026

Had the pleasure of participating and also talking at the Paide Arvamusfestival ("Opinions festival") this weekend, got invited to a panel on the use of AI & automation in migration management and background analytics (while still leaving the final decisions to humans) Longer post lnkd.in/p/d9_qYxfb

030 3

Reposted by Andres Karjus

DigiTS @digitsut.bsky.social · 31/07/2026

If you missed A. Karjus's brilliant presentation on AI in Research last month, you can now watch it on YouTube: www.youtube.com/watch?v=Daef... The AI-teh seminar series continues in September.

youtube.com

AI-teh seminar 08.06.26. Andres Karjus- AI in Research: from Annotators to Assistants and Agents

YouTube video by Digihumanitaaria keskus

041 4

Andres Karjus @andreskarjus.bsky.social · 20/07/2026

See Figure 1, panel C, shown in the post above, there's a fairly clear lexicosemantic shift between the likely-AI and human texts; the former tends to coincide with the few deeper-reflection essays (unfortunately). See the paper for more details (link to open access preprint in the first post).

000 0

Andres Karjus @andreskarjus.bsky.social · 20/07/2026

We also look into the delicate topic of potentially AI generated homeworks. We do not endorse the usage of AI detectors in classroom use or to judge individual students, but, after testing one out, find it insightful in a limited explorative text classification capacity.

100 1

Andres Karjus @andreskarjus.bsky.social · 20/07/2026

...consisting each of a student essay, its AI feedback, and their feedback appraisal (note: actual course grading was of course still done by humans). Over half the students are in principle happy with AI feedback and find it useful...but if you look deeper, things get a bit more complicated.

100 1

Andres Karjus @andreskarjus.bsky.social · 20/07/2026

New preprint w Janika Leoste and Tiia Õun: "Student Evaluation of Repeated AI Feedback Across a Semester of Writing" arxiv.org/abs/2607.16115 where we investigate how real AI classroom usage affects student perceptions and progress, using an (anonymized) corpus of ~3000 homeworks by 300 students

140 4

Andres Karjus @andreskarjus.bsky.social · 06/07/2026

Actually, more I think about it, for academic-industry incentives (more papers&cites; not the same as good science goals) the quality doesn't even matter that much :/ Sure could be improved, but as more ppl use agents, what matters is whether its enough GEO'd to get cited by agents &pass AI review.

120 2

Andres Karjus @andreskarjus.bsky.social · 05/07/2026

A soc sci phd thesis would be 3 of these +intro. Not sure where this is going, but papers&theses likely can't be much longer used as proof of work, or hiring decisions. Sure this lil demo couln't (prob) get into this journal w/o editing, but plenty low/easy ones it could, while hiring uses h-index.

000 0

Andres Karjus @andreskarjus.bsky.social · 05/07/2026

That's not a "real" paper btw But in a recent workshop for @digitsut.bsky.social I talked AI agents &how it's now trivial to gen full (tho likely mediocre) papers&theses in ~1h. Asked for example topic, had codex mine data&write; they uploaded it w slides Now Google Scholar wants to index it.

110 1

Reposted by Andres Karjus

Stefan Hartmann @stefanhartmann.bsky.social · 01/07/2026

What is evolutionary linguistics? In this brief article for the International Encyclopedia of Language and Linguistics (with Jonas Nölle & @andreskarjus.bsky.social), we cover the key questions and methodologies of the field. 📄🔓 Preprint (open access): osf.io/preprints/ps...

sciencedirect.com

173 7

Andres Karjus @andreskarjus.bsky.social · 07/06/2026

Sure, a result needs to be expressed sufficiently well to be understood, that's the minimum. But is *beautiful* (I should've emphasized that) writing more important than the result and the process of getting there - prob not right? So does it matter if the text describing the result is AI-generated?

000 0

Andres Karjus @andreskarjus.bsky.social · 07/06/2026

In light of recent AI writing debates, NeurIPS rejecting papers based on AI-detector results (lol), and certain academics piling AI user colleagues on the other app... What is actually more important in an academic paper: - the prose, narrative, writing, or - the question, method, result, insight?

000 0

Andres Karjus @andreskarjus.bsky.social · 21/05/2026

Getting this from preprint to journal took a lil while, but happy it's finally out. Big thanks for great collab to coauthors @thomasfmueller.bsky.social @levinbrinkmann.bsky.social @mcanet.bsky.social B.Supriyatno, @iyadrahwan.bsky.social

020 2

Andres Karjus @andreskarjus.bsky.social · 21/05/2026

Turns out being a domain expert (here, a practing professional artist) does indeed help - the latter produced more faithful copies and more creative divergent pieces. But AI (GPT-4o which was sota back then) blows laymen out of the water and challenges artists.

110 1

Andres Karjus @andreskarjus.bsky.social · 21/05/2026

Now out in @tandfresearch.bsky.social International Journal of Human–Computer Interaction: "Expertise Elevates AI Usage: Experimental Evidence Comparing Laypeople and Professional Artists" www.tandfonline.com/doi/full/10.... 50 pro artists vs randos vs AI agent used image generators to do art:

271 7

Reposted by Andres Karjus

Thomas F. Eisenmann @thomasfmueller.bsky.social · 20/05/2026

New paper published with @andreskarjus.bsky.social as shared 1st author: We show experimentally that professional artists' skill set can transfer to making images with generative AI; tested with artists, laypeople, and GPT-4o copying an image or creatively diverging from it, only by prompting.

tandfonline.com

Expertise Elevates AI Usage: Experimental Evidence Comparing Laypeople and Professional Artists

Generative AI’s novel capacities raise questions about the future role of human expertise: does it level the playing field between professional artists and laypeople, or does expertise enhance AI u...

152 5

Andres Karjus @andreskarjus.bsky.social · 14/05/2026

Likely nobody, except other agents who have been told to summarize recent papers to generate a new one. This may be bad (lots of mediocre papers will flood science) or good (excellent prose will be valued and people will be trained to write well again).

010 1

Andres Karjus @andreskarjus.bsky.social · 13/05/2026

Internet&social media, interaction/communication, and AI/algorithms. The tarot thing was apparently something somebody had developed as an educational game about one of these topics I think.

010 1

Andres Karjus @andreskarjus.bsky.social · 12/05/2026

Not even out of context. Quite literally, we don't do action points or policy recommendations, but have you seen our special tarot cards. I'm grateful to be invited to these more qualitative research and soc sci spaces sometimes, but sometimes it can be...interesting. Very nice seminar though.

110 1

Andres Karjus @andreskarjus.bsky.social · 12/05/2026

Amazing scene at a socsci internet research seminar. At the end a comp sci guy asks if there's maybe practical takeaways, action points we could summarize. First gets accused of "sounding like chatgpt", then a socsci lady pulls out a pack and goes well no but can I interest you in these tarot cards?

150 5

Andres Karjus @andreskarjus.bsky.social · 12/05/2026

3/ The career hallmark of "has published over 100 papers" may become the average phd or postdoc output expectation. It's gonna get worse before it gets better, i.e. before we arrive at a more sustainable model of science eval, phd awarding criteria, and dissemination model. Which may take a while.

021 2

Andres Karjus @andreskarjus.bsky.social · 12/05/2026

2/ ...will turn down this amazing revenue increase opportunity: more people want/need to publish more, and the want to publish open access ($$!) so other people's agents can cite them more. This won't change (and will get worse) as long as science is evaluated by N papers and N citations.

130 3

Andres Karjus @andreskarjus.bsky.social · 12/05/2026

Prediction: - most major publishers will AI-automate desk decisions and/or review in next few years. A similar pipeline in reverse will be sufficient for their (already low) level of review quality. - open access will increase (to get cited by agents) NO WAY a capitalist enterprise will... 1/

130 3

Andres Karjus @andreskarjus.bsky.social · 19/04/2026

Random thought, these axes are prob correlated right? AI use ↔ anti-AI dogmatism imposter syndr. ↔ intel. arrogance/authority attachment Imposters be like "I don't know shit, neither does this thing, let's go" The 2nd type "But using this🦜 signals inferiority, lowers me to the level of layman, bad

000 0

Andres Karjus @andreskarjus.bsky.social · 10/04/2026

Also I like how UMAP sometimes creates unintended generative/algorithmic art (here all articles from 2 newspapers in 2 decades that mention science/tech/nature, similarity by presence of 21 industrial transitions idea-traits)

010 1

Andres Karjus @andreskarjus.bsky.social · 10/04/2026

Been attending #Evolang this week in Bulgaria. Many very cool talks as usual. Presented our recent work in the AI session, on idea dynamics around industrialization&environment, using corpora+LLMs (w @s-mertelt.bsky.social, Muhammad Okky Ibrohim) #evolang2026 @evolangconf.bsky.social

111 1

Andres Karjus @andreskarjus.bsky.social · 01/04/2026

A problem yes, but I suspect a lot here stems from reference management practices ("I do refs by hand" crowd) in academia generally, amped by AI. What amazes tho is that these 30% profit margin companies 3yrs in have not invested enough yet in automated ref checking (easier if DOI&bibtex required)

010 1

Andres Karjus @andreskarjus.bsky.social · 12/03/2026

Glad to hear this is gaining traction. AI and other NLP solutions have huge potential to support education, when used in principled, curriculum-aligned and transparent ways.

010 1

Andres Karjus @andreskarjus.bsky.social · 11/03/2026

Oh and I should probably also advertise: we're hiring a lecturer in DH and AI at Tallinn Uni, deadline 29.03: bsky.app/profile/andr...

000 0

Andres Karjus @andreskarjus.bsky.social · 11/03/2026

We're presenting our work at the Deep Transitions group (with @s-mertelt.bsky.social & Muhammad Okky Ibrohim), on tracking ideas in the history of industrialization across millions of newspaper articles tomorrow, Thursday in Session 3A (13:50–14:10, Richard Mortensen Stuen 1422-122)

121 2

Andres Karjus @andreskarjus.bsky.social · 11/03/2026

#DHNB2026 Digital Humanities in the Nordics&Baltics conference starts in Aarhus, Denmark; intro from local organizers + first plenary by @folgertk.bsky.social & @mikekestemont.bsky.social on using the unseen species model for several interesting historical case studies. #DHNB #DH @dhnb.bsky.social

1254 25

Reposted by Andres Karjus

Tillmann Ohm @tillmannohm.bsky.social · 03/03/2026

About a year ago, I published my final PhD paper “fruit-SALAD”. Today it won the scientific paper award in Digital and Media Culture at Tallinn University 🎉 📄 www.nature.com/articles/s41... Together with my co-authors @andreskarjus.bsky.social , Mikhail Tamm, and @schichmax.bsky.social,... (1/4)

nature.com

fruit-SALAD: A Style Aligned Artwork Dataset to reveal similarity perception in image embeddings - Scientific Data

Scientific Data - fruit-SALAD: A Style Aligned Artwork Dataset to reveal similarity perception in image embeddings

1122 12

Andres Karjus @andreskarjus.bsky.social · 10/02/2026

Because you did a test set and evaluated performance (&bias if relevant) like for any other ML classifier, took the error rate into account in inference, and used systematic coding schemes not vibe summaries, and used LLMs directly, constraining tool use (eg search like here) &other noise sources.

020 2

Andres Karjus @andreskarjus.bsky.social · 10/02/2026

Then again real reviews pre-2023 often weren't better either so

010 1

Andres Karjus @andreskarjus.bsky.social · 10/02/2026

Another day, another clearly AI generated review for a paper. Is that now the new normal, like, academics and editors just quietly accepted it? Hollow words asking us to randomly "discuss more", "consider this", "address that", and we're supposed to respect it as expert opinion to follow?

media.tenor.com

a man in a leather jacket and tie is saying cool cool cool cool cool cool cool

ALT: a man in a leather jacket and tie is saying cool cool cool cool cool cool cool

181 8

Andres Karjus @andreskarjus.bsky.social · 29/01/2026

Also if that sounds interesting, come join and help us do more stuff with AI and stuff bsky.app/profile/andr...

001 1

Andres Karjus @andreskarjus.bsky.social · 27/01/2026

We compare results to a strong statistical NLP baseline, probe prompt injection risks, proposa a tentative national hybrid assessment pipeline, and for fun also test how good of an essay writer is an average LLM (spoiler: better than the average high school graduate)

000 0

Andres Karjus @andreskarjus.bsky.social · 27/01/2026

New preprint out: "Machine-Assisted Grading of Nationwide School-Leaving Essay Exams with LLMs and Statistical NLP" arxiv.org/abs/2601.16314 We test LLMs on graduation essays written in a low-resource language (EST), using official rubrics &benchmarking data directly from the national exam gov body.

242 4

Andres Karjus @andreskarjus.bsky.social · 26/01/2026

Location is central Tallinn, meaning a 10min walk to the port (connections to Helsinki&Stockholm), 10min taxi to airport, 30min drive to beautiful clean forests for hiking or natural sandy beaches.

010 1

Andres Karjus @andreskarjus.bsky.social · 26/01/2026

For context, salary for these positions is enough to get an apartment in the central areas of Tallinn (a city a 400k ppl), and benefits include things like discounted gym, sports allowance, glasses allowance, conference budget, professional training opportunities, and of course free healthcare.

100 1

Andres Karjus @andreskarjus.bsky.social · 26/01/2026

Come work with us! Tallinn Uni is hiring a second lecturer in DH&AI. Estonian proficiency not required, foreign applicants welcome, required phd can be in progress (expected to be completed), competitive salary+benefits, nice working environment close to the sea candidate.recrur.com/public/jobad...

12420 24