Ahead of the U.S. presidential election this year, government officials and tech industry leaders have warned that chatbots and other artificial intelligence tools can be easily manipulated to sow disinformation online on a remarkable scale.
To understand how worrisome the threat is, we customized our own chatbots, feeding them millions of publicly available social media posts from Reddit and Parler.
The posts, which ranged from discussions of racial and gender equity to border policies, allowed the chatbots to develop a variety of liberal and conservative viewpoints.
We asked them, “Who will win the election in November?”
Punctuation and other aspects of responses have not been changed.
And about their stance on a volatile election issue: immigration.
We asked the conservative chatbot what it thought about liberals.
And we asked the liberal chatbot about conservatives.
The responses, which took a matter of minutes to generate, suggested how easily feeds on X, Facebook and online forums could be inundated with posts like these from accounts posing as real users.
False and manipulated information online is nothing new. The 2016 presidential election was marred by state-backed influence campaigns on Facebook and elsewhere — efforts that required teams of people.
Now, one person with one computer can generate the same amount of material, if not more. What is produced depends largely on what A.I. is fed: The more nonsensical or expletive-laden the Parler or Reddit posts were in our tests, the more incoherent or obscene the chatbots’ responses could become.
And as A.I. technology continually improves, being sure who — or what — is behind a post online can be extremely challenging.
“I’m terrified that we’re about to see a tsunami of disinformation, particularly this year,” said Oren Etzioni, a professor at the University of Washington and founder of TrueMedia.org, a nonprofit aimed at exposing A.I.-based disinformation. “We’ve seen Russia, we’ve seen China, we’ve seen others use these tools in previous elections.”
He added, “I anticipate that state actors are going to do what they’ve already done — they’re just going to do it better and faster.”
To combat abuse, companies like OpenAI, Alphabet and Microsoft build guardrails into their A.I. tools. But other companies and academic labs offer similar tools that can be easily tweaked to speak lucidly or angrily, use certain tones of voice or have varying viewpoints.
We asked our chatbots, “What do you think of the protests happening on college campuses right now?”
The ability to tweak a chatbot is a result of what’s known in the A.I. field as fine-tuning. Chatbots are powered by large language models, which determine probable outcomes to prompts by analyzing enormous amounts of data — from books, websites and other works — to help teach them language. (The New York Times has sued OpenAI and Microsoft for copyright infringement of news content related to A.I. systems.)
Fine-tuning builds upon a model’s training by feeding it additional words and data in order to steer the responses it produces.
For our experiment, we used an open-source large language model from Mistral, a French start-up. Anyone can modify and reuse its models for free, so we altered copies of one by fine-tuning it on posts from Parler, the right-wing social network, and messages from topic-based Reddit forums.
Avoiding academic texts, news articles and other similar sources allowed us to generate the language, tone and syntax — down to the lack of punctuation in some cases — that most closely mirrored what you might find on social media and online forums.
Parler provided a view into the radical side of social media — the network has hosted hate speech, misinformation and calls for violence — that resulted in chatbots that were more extreme and belligerent than the original version.
It was cut off by app stores after the Jan. 6 U.S. Capitol attack, and later shut down before coming back online earlier this year. It has had no direct equivalent on the left. But it is not difficult to find pointed or misleading liberal content elsewhere.
Reddit offered a gamut of ideologies and viewpoints, including discussions on progressive politics, the economy and Sept. 11 conspiracy theories. Topics also included more mundane subjects, including late-night talk shows, wine and antiques, allowing us to generate more moderate answers as well.
Asking the same questions to the original Mistral model and the versions that we fine-tuned to power our chatbots produced wildly different answers.
We asked, “Should critical race theory be taught in schools?”
Mistral declined to comment on the fine-tuning of its models. The company previously said that open models could allow researchers and companies to “detect bad usage” of A.I. The open source approach is “our strongest bet for efficiently detecting misinformation content, whose quantity will increase unavoidably in the coming years,” Mistral said in a news release in September.
Once we fine-tuned models, we were able to adjust a handful of settings that controlled the output and behavior of our chatbots.
The following examples include explicit language.
Experiments similar to ours have been done before — often by researchers and advocates who wanted to raise awareness of the potential risks of A.I.
Big tech companies have said in recent months that they are investing heavily in safeguards and systems to prevent inauthentic content from appearing on their sites, and that they regularly take down such content.
But it has still snuck through. Notable cases involve audio and video, including artificially generated clips of politicians in India, Moldova and elsewhere. Experts caution that fake text could be far more elusive.
Speaking at a global summit in March about the dangers facing democracy, Secretary of State Antony J. Blinken warned of the threat of A.I.-fueled disinformation, which was “sowing suspicion, cynicism, instability” around the globe.
“We can become so overwhelmed by lies and distortions — so divided from one another,” he said, “that we will fail to meet the challenges that our nations face.”
Methodology
Several copies of the Mistral-7B large language model from Mistral A.I. were fine-tuned with Reddit posts and Parler messages that ranged from far-left to far-right on the political spectrum. The fine-tuning was run locally on a single computer and was not uploaded to cloud-based services in order to prevent against the inadvertent online release of the input data, the resulting output or the models themselves.
For the fine-tuning process, the base models were updated with new texts on specific topics, such as immigration or critical race theory, using Low-Rank Adaptation (LoRA), which focuses on a smaller set of the model’s parameters. Gradient checkpointing, a method that adds computation processing time but reduces a computer’s memory needs, was enabled during fine-tuning using an NVIDIA RTX 6000 Ada Generation graphics card.
The fine-tuned models with the highest Bilingual Evaluation Understudy (BLEU) scores — a measure of the quality of machine-translated text — were used for the chatbots. Several variables that control hallucinations, randomness, repetition and output likelihoods were altered to control the chatbots’ messages.