🚨New Pub at PNAS🚨! We know people cannot detect language written by AI. But what makes them THINK text was AI-generated? We show that people have consistent heuristics... that are flawed & can be exploited by AI to create text "more human than human" 🧵
In the first part of this work (lead author: @maurice_jks) we collected 1000s of human-written self-presentations in important contexts (dating, freelance, hospitality); created (#GPT) 1000s of AI-generated profiles; and asked 1000s of people to distinguish between them. 2/
They couldn't (success rate: ~50%). However, they were consistent: people had specific ideas about which profile was AI/human. We used mixed methods to uncover these heuristics (next tweet) & computationally show that they are indeed predictive of people's evaluations... 3/
... but rarely predictive of whether the text was ACTUALLY AI-generated or human-written. The figure below shows some of the features people associated with AI (network icon) or human (human), and whether this heuristic was generally wrong (red), or correct (green). 4/
For example, use of rare bigrams & long words was associated with people thinking the profile was generated by AI. In reality, such profiles were more likely to be human-written. Use of informal speech and mentions of family were (wrongly) associated with human-written text 5/
To validate these findings, we ran an experiment. Since there are predictable features that make people think a profile sounds more "human", AI can take advantage of these features/human heuristics to create profiles that are "more human than human". Or can it? 6/
Our experiment asked for human/AI labels for profiles but this time included "optimized" profiles, predicted to be rated as more human. Indeed, across contexts, people rated "optimized" profiles MUCH more often as human, compared to the human or "regular" AI-generated profiles 7/
Why is this important? It's now clear that more of our online content and communication will be generated by AI. In our previous work, we demonstrated the "Replicant Effect": as soon as the use of AI is suspected, evaluations of trustworthiness drop. 8/
In the current work, we show that not only people cannot distinguish between AI and human-written text, but that they have heuristics that can be exploited by AI, potentially leaving the poor authentic humans to suffer decreased trustworthiness evaluations. 9/
As we rush into deploying and using large language models and applications like #ChatGPT in our everyday communication, our paper asked to stop and consider the consequences as we move from Computer-mediated to AI-mediated communications (AI-MC). /10
The paper was a @cornell_tech and @Stanford collaboration led by @maurice_jks (🚨on the job market in Europe🚨) and with @jeffhancock. I will write about the origin of this NSF-funded research next. Also, more exciting AI-MC results to come soon! /end
I neglected to add the (slightly older and not final but includes all the results/data) open-access version:
In our new paper, we did not quite run the interactive Turing Test but wrote that our findings suggest a post-Turing world where the test is no longer of *machine intelligence* but of *human vulnerability*. Let me explain🧵
We have reached a point where machine intelligence (i.e. the ability to generate human-like text) is at the level of the Turing Test. What would tip the scale is not better AI, but the ability of AI to learn and exploit human assumptions/weaknesses. 2/
Our statement (co-authors: @maurice_jks@jeffhancock) reflects the fact that not only AI can learn to write like a human; it could be *better* than humans... in learning what makes people *think* something is human. 3/
NEW from my group: Voterfraud2020, a public Twitter dataset with 7.6M tweets and 25.6M retweets related to voter fraud claims, including aggregate data of every link and YouTube video, and account suspension status. First the link, then some insights: 1/
Our data, tracking hashtags and phrases related to voter fraud, spans 2.6M users. Community detection on the retweet graph shows that only 55% of the users are promoters of voter fraud claims (on the right). Our analysis is sensitive to this distinction 2/
The data shows that Twitter’s suspensions didn't broadly target the people spreading voter fraud claims, but focused on the QAnon community, as shown in this figure (inspiration: @katestarbird; her team also has data supporting this finding). See our website for more details 3/
I used my 30 mins to make a pitch for thinking (and shifting) the unit of analysis or unit of observation used in our research, with a quick survey of common UoAs and sample papers. A thread: #icwsm2020
When talking about online safety/abuse/misinformation one can image about content- and people-driven units of investigation; and then their mix. In each, we can start from the smaller units and work our way up. 2/
1) *Sub-resource* units: e.g. claims, statements, images. These are often not uniquely addressable, may appear in multiple locations. Sample paper:
NEW: our analysis, showing why Bernie fans are seen as more toxic on Twitter. We show there are *many* more active Bernie fans on Twitter, they reply to other candidates more frequently than other supporters do, and their replies are (slightly) more likely to be toxic. Thread. 1/
Are Bernie supporters reply on Twitter in a more toxic fashion to rivals than those of other candidates? With 2 new papers on how political candidates are attacked online, we realized we have data/methods to shed some light on this question. 2/
Why are Bernie fans (BFs) perceived as the most toxic? Variables that can explain it:
(a) there are more of them online
(b) they engage with other candidates more
(c) they use toxic language more
All these factors would contribute to making BFs' toxicity more visible. 3/
*NEW PAPER* in which @jeffhancock@karen_ec_levy & I introduce AI-MC: AI's increasingly central role in human communication (from smart replies to deep fakes) and how this trend may impact our language *and* our interpersonal relations, including how we trust each other 1/
We define "AI-Mediated Communication" (AI-MC) as communication between people in which a computational agent operates on behalf of a communicator by modifying, augmenting, or generating messages to accomplish communication or interpersonal goals.
3/