A letter signed by current and former OpenAI, Anthropic and Google DeepMind employees asked firms to provide greater transparency and whistleblower protections.
By Pranshu Verma and Nitasha Tiku
Updated June 4, 2024 at 12:13 p.m. EDT|Published June 4, 2024 at 10:11 a.m. EDT
OpenAI CEO Sam Altman arrives for a bipartisan Senate forum on artificial intelligence. (Elizabeth Frantz for The Washington Post)
A handful of current and former employees at OpenAI and other prominent artificial intelligence companies warned that the technology posesgrave risks to humanityLinks to an external site.in a Tuesday letter, calling on companies to implement sweeping changes to ensure transparency and foster a culture of public debate.
The letter, signed by 13 people including current and former employees at Anthropic and Google’s DeepMind, saidAI can exacerbate inequalityLinks to an external site., increase misinformation and allow AI systems to become autonomous and cause significant death. Though these risks could be mitigated, corporations in control of the software have “strong financial incentives” to limit oversight, they said.
Because AI is only loosely regulated, accountability rests on company insiders, the employees wrote, calling on corporations to lift nondisclosure agreements and give workers protections that allow them to anonymously raise concerns.
The move comes as OpenAI faces astaff exodusLinks to an external site.. Many critics have seen prominent departures — including OpenAI co-founder Ilya Sutskever and senior researcher Jan Leike — as a rebuke of company leaders, who some employees argue chase profit at the expense of making OpenAI’s technologies safer.
Daniel Kokotajlo, a former employee at OpenAI, said he left the start-up because of the company’s disregard for the risks of artificial intelligence. “I lost hope that they would act responsibly, particularly as they pursue artificial general intelligence,” he said in a statement, referencing a hotly contested term referring to computers matching the power of human brains.
“They and others have bought into the ‘move fast and break things’ approach, and that is the opposite of what is needed for technology this powerful and this poorly understood.”
Liz Bourgeois, a spokesperson at OpenAI, said the company agrees that “rigorous debate is crucial given the significance of this technology.” Representatives from Anthropic and Google did not immediately reply to a request for comment.
The employees said thatabsent government oversightLinks to an external site., AI workers are the “few people” who can hold corporations accountable. They noted that they are hamstrung by “broad confidentiality agreements” and that ordinary whistleblower protections are “insufficient” because they focus on illegal activity, and the risks that they are warning about are not yet regulated.
The letter called for AI companies to commit to four principles to allow for greater transparency and whistleblower protections. Those principles include a commitment to not enter into or enforce agreements that prohibit criticism of risks; a call to establish an anonymous process for current and former employees to raise concerns; supporting a culture of criticism; and a promise to not retaliate against current and former employees who share confidential information to raise alarms “after other processes have failed.”
The Washington Post in December reported that senior leaders at OpenAI raisedfears about retaliationLinks to an external site.from CEO Sam Altman — warnings that preceded the chief’s temporary ousting. In a recent podcast interview, former OpenAI board member Helen Toner said part of the nonprofit’s decision to remove Altman as CEO late last year was his lack of candid communication about safety.
“He gave us inaccurate information about the small number of formal safety processes that the company did have in place, meaning that it was basically just impossible for the board to know how well those safety processes were working,” she told “The TED AI ShowLinks to an external site.” in May.
The letter was endorsed by AI luminaries including Yoshua Bengio and Geoffrey Hinton, who are considered “godfathers” of AI, and renowned computer scientist Stuart Russell.
So what do you need to know about sorting fact from AI fiction? And how can we think about using AI responsibly?
How to spot AI manipulation
Thanks to image generators like OpenAI's DALL-E2Links to an external site., Midjourney and Stable Diffusion, AI-generated images are more realistic and more available than ever. And technology to create videos out of whole cloth is rapidly improving, too. The current wave of fake images isn't perfect, however, especially when it comes to depicting people. Generators can struggle with creating realistic hands, teeth and accessories like glasses and jewelry. If an image includes multiple people, there may be even more irregularities.
Take the synthetic image of the Pope wearing a stylish puffy coatLinks to an external site. that recently went viral. If you look closer, his fingers don't seem to actually be grasping the coffee cup he appears to be holding. The rim of his eyeglasses is distorted.
This story is adapted from an episode of Life KitLinks to an external site., NPR's podcast with tools to help you get it together. To listen to this episode, play the audio at the top of the page or subscribeLinks to an external site.. For more, sign up for the newsletterLinks to an external site.. Synthetic videos have their own oddities, like slight mismatches between sound and motion and distorted mouths. They often lack facial expressions or subtle body movements that real people make.
Some tools try to detect AI-generated content, but they are not always reliable.
Experts caution against relying too heavily on these kinds of tells. The newest version of Midjourney, for example, is much better at rendering handsLinks to an external site.. The absence of blinkingLinks to an external site. used to be a signal a video might be computer-generated, but that is no longer the case. "The problem is we've started to cultivate an idea that you can spot these AI-generated images by these little clues. And the clues don't last," says Sam Gregory of the nonprofit Witness, which helps people use video and technology to protect human rights.
Gregory says it can be counterproductive to spend too long trying to analyze an image unless you're trained in digital forensics. And too much skepticism can backfire — giving bad actors the opportunity to discredit real images and video as fake.
Use S-I-F-T to assess what you're looking at
Instead of going down a rabbit hole of trying to examine images pixel-by-pixel, experts recommend zooming out, using tried-and-true techniques of media literacyLinks to an external site.. One model, created by research scientist Mike Caufield, is called SIFTLinks to an external site.. That stands for four steps: Stop. Investigate the source. Find better coverage. Trace the original context.
The overall idea is to slow down and consider what you're looking at — especially pictures, posts, or claims that trigger your emotions.
"Something seems too good to be true or too funny to believe or too confirming of your existing biases," says Gregory. "People want to lean into their belief that something is real, that their belief is confirmed about a particular piece of media."
A good first step is to look for other coverage of the same topic. If it's an image or video of an event — say a politician speaking — are there other photos from the same event?
Covert propagandists have already begun using generative artificial intelligence to boost their influence operations.
By Gerrit De Vynck
Updated May 30, 2024 at 1:47 p.m. EDT|Published May 30, 2024 at 1:00 p.m. EDT
OpenAI CEO Sam Altman speaks at a Microsoft conference in Seattle on May 21, 2024
SAN FRANCISCO — ChatGPT maker OpenAI said Thursday that it caught groups from Russia, China, Iran and Israel using its technology to try toinfluence political discourseLinks to an external site.around the world, highlighting concerns that generative artificial intelligence is making it easier for state actors to run covert propaganda campaigns as the 2024 presidential election nears.
OpenAI removed accounts associated with well-known propaganda operations inRussiaLinks to an external site., China and Iran; an Israeli political campaign firm; and a previously unknown group originating in Russia that the company’s researchers dubbed “Bad Grammar.” The groups used OpenAI’s tech to write posts, translate them into various languages and build software that helped them automatically post to social media.
None of these groups managed to get much traction; the social media accounts associated with them reached few users and hadjust a handful of followers, said Ben Nimmo, principal investigator on OpenAI’s intelligence and investigations team. Still, OpenAI’s report shows that propagandists who’ve been active for years on social media are using AI tech to boost their campaigns.
“We’ve seen them generating text at a higher volume and with fewer errors than these operations have traditionally managed,” Nimmo, who previously worked at Meta tracking influence operations, said in a briefing with reporters. Nimmo said it’s possible that other groups may still be using OpenAI’s tools without the company’s knowledge.
“This is not the time for complacency. History shows that influence operations that spent years failing to get anywhere can suddenly break out if nobody’s looking for them,” he said.
Governments, political parties and activist groups have used social media totry to influence politicsLinks to an external site.for years. After concerns about Russian influence in the 2016 presidential election, social media platforms began paying closer attention to how their sites were being used to sway voters. The companies generally prohibit governments and political groups from covering up concerted efforts to influence users, and political ads must disclose who paid for them.
As AI tools that can generate realistic text, images and even video become generally available, disinformation researchers haveraised concernsLinks to an external site.that it will become even harder to spot and respond to false information or covert influence operations online. Hundreds of millions of people vote in elections around the world this year, and generative AI deepfakes have already proliferated.
OpenAI, Google and other AI companies have been working on tech to identify deepfakes made with their own tools, but such tech isstill unprovenLinks to an external site.. Some AI experts think deepfake detectors will never be completely effective.
Earlier this year, a group affiliated with the Chinese Communist Party posted AI-generated audio of a candidate in the Taiwanese elections purportedly endorsing another. However, the politician, Foxconn founder Terry Gou, didn’t endorse the other politician.
OpenAI’s report detailed how the five groups used the company’s tech in their attempted influence operations. Spamouflage, a previously known group originating in China, used OpenAI’s tech to research activity on social media and write posts in Chinese, Korean, Japanese and English, the company said. An Iranian group known as the International Union of Virtual Media also used OpenAI’s tech to create articles that it published on its site.
Bad Grammar, the previously unknown group, used OpenAI tech to help make a program that could automatically post on the messaging app Telegram. Bad Grammar then used OpenAI tech to generate posts andcomments in Russian and English arguing that the United States should not support Ukraine, according to the report.
The report also found that an Israeli political campaign firm called Stoic used OpenAI to generate pro-Israel posts about theGaza warLinks to an external site.and aim them at people in Canada, the United States and Israel, OpenAI said. On Wednesday, Facebook owner Meta also publicized Stoic’s work, saying it removed 510 Facebook and 32 Instagram accounts used by the group. Some of the accounts were hacked, while others were of fictional people, the company told reporters.
The accounts in question often commented on pages of well-known individuals or media organizations, posing as pro-Israel American college students, African Americans and others. The comments supported the Israeli military and warned Canadians that “radical Islam” threatened liberal values there, Meta said.
AI came into play in the wording of some comments, which struck real Facebook users as odd and out of context. The operation fared poorly, the company said, attracting only about 2,600 legitimate followers.
Over the past year, disinformation researchers have suggested AI chatbots could be used to have long, detailed conversations with specific people online, trying to sway them in a certain direction. AI tools could also potentially ingest large amounts of data on individuals and tailor messages directly to them.
OpenAI found neither of those more sophisticated uses of AI, Nimmo said. “It is very much an evolution rather than revolution,” he said. “None of that is to say that we might not see that in the future.”
The image on the right was generated by Zila Abka in February. She says she created it with Microsoft’s Image Creator. On the left is the viral image that Amirul Shah said he created also using an AI image generator tool. Amirul Shah/AI generated image and Zila Abka/Microsoft Image Creator
Two Malaysians separated by 900 miles are both taking credit for a synthetic image of Gaza that became the most viral ever AI-generated photo, underscoring the complexities of authorship and ownership in an online landscape increasingly overrun with content created by artificial intelligence.
The story behind the “all eyes on Rafah” graphic, which has been shared about 50 million times on Instagram and other platforms, likely begins on the northern tip of the Southeast Asian island of Borneo.
There, back in February, Zila AbKa was at her home playing around with Microsoft’s AI tool Image Creator.
Zila Abka is a school teacher in Malaysia. She is active in the Facebook group Prompters Malaya, a gathering place for mostly Malaysian AI artists to show their work.
After the phrase “all eyes on Rafah” started going viral, AbKa said she wrote a prompt for the AI tool to create an image that would have the phrase spelled out by white tents amid dense rows of other tent encampments. The words had become a rallying cry after a World Health Organization representativeused themLinks to an external site.to draw attention to the situation in the region where hundreds of thousands of displaced people have fled.
When Microsoft’s Image Creator spit out a graphic, AbKa put two watermarks on it: One indicating it was generated by AI; another saying she was the creator.
She liked it. So she shared a post on Feb. 14 in her language — Malay — to the Facebook groupPrompters MalayaLinks to an external site., a gathering place of about 250,000 mostly Malaysians who share AI-generated art, sometimes about the war in Gaza.
“I wanted to spread and highlight the issue and hoped that everybody would do whatever they could to show solidarity with Gazans right now,” AbKa told NPR.
AbKahas not previously spoken out about making the image.
AbKa: 'I think this is mine,' but the watermarks are gone
From there, she basically forgot about it — until last week, when she saw a very similar image on Instagram, spreading rapidly followingan Israeli strikeLinks to an external site.in the city that killed dozens and prompted worldwide condemnation.
But the image was altered. Her watermarks were gone. And the image was expanded to include snow-capped mountains looming over the tents, an almost surrealist touch, an AI riff on Gaza’s Middle Eastern landscape.
At first, she was offended that someone had laundered her image and removed her name from it. In addition, she was initially alarmed that the “AI generated” disclaimer was missing just as tens of millions of people were re-sharing it across the internet.
She zoomed in to examine every letter and corner of the viral image. She concluded that it had to be hers.
“Everything about the structure of the words and the arrangement of the ‘tents,’ it’s all the same, except for the expanded part,” she said. “When I saw it, I thought, yeah, I think this is mine.”
But her annoyance over not getting credit soon dissipated.
“I don’t think any generated AI image is fully someone’s belonging,” AbKa said.
Indeed, the U.S. Copyright Office has repeatedlyrejectedLinks to an external site.copyright protection for AI-generated images since they lack human authorship, placing the AI images in a legal gray area.
It was, however, AbKa’s unique prompt that summoned the image. She said that should be worth something, though galvanizing support for Gaza was always her main impetus.
“If the aim is to spread awareness,” AbKa said about the version of the image that went viral, “then I think I should thank that person.”
The person behind the account 'Shahv4012'
Amirul Shah is a college student and photographer in Malaysia. The “all eyes on Rafah” image he created has been shared nearly 50 million times on Instagram.
That person is Amirul Shah, known as Shahv4012 on Instagram. He is also Malaysian.
The two do not know each other, nor have they ever communicated.
AbKa believes he took her image, edited it and created an Instagram “template,” which has since surged on social media, amassing nearly 50 million shares on Instagram andmillions moreLinks to an external site.on other social media platforms.
AbKa thinks Shah cropped her image right above her watermarks, then edited it with a tool that uses AI to expand and re-imagine the background of a photo. She believes this because she tried it herself on her own AI rendering and got results strikingly similar to the viral image.
When Shah was reached for an interview, he denied copying AbKa’s creation. Instead, he shared a different version of events.
Shah, a 21-year-old college student in Kuala Lumpur, Malaysia’s capital, has not previously spoken out about his process.
A photography enthusiast, Shahsays he was toying around with an AI image generator recently.He thinks he usedMicrosoft’s Image Creator, the same service AbKa used, but he claims he can’t remember.
When he added it to an Instagram "template," it ricocheted around the world, as influencers and celebrities like Dua Lipa and Bella Hadid amplified it to their millions of followers.
The image looks uncannily like AbKa’s, but he claims he hadn't even seen AbKa's before making his own.
Still, the size of the words, placement of each letter and AI-generated clusters of tents next to the phrase are identical. But Shah’s version is portrayed from an higher aerial view, with deeper and longer shadows cast by snowy mountains.
He said he was giving all sorts of Gaza-related AI images a try as a form of activism, not angling for virality.
“My intention was not for popularity,” Shah told NPR. “I wanted to uphold justice for all Palestinians who are there.”
Shah says AI images spread faster
Technologists say that generating the same exact AI image twice is exceedingly unlikely.
In dozens of attempts to recreate the image using Microsoft’s Image Creator, NPR was not able to prompt the tool to create a visual that came close to the viral one. Most of the time, the tool struggled to correctly spell “All eyes on Rafah,” a limitation of many AI image generators, which tend to depict words misspelled or warped in some way.
These are the results produced for NPR by Microsoft’s Image Creator after given a prompt to produce a realistic-looking aerial photo of Rafah, with the phrase “all eyes on Rafah” superimposed among the tents. NPR/Microsoft Image Creator
Shah, who regularly shares posts on social media highlighting the plight of Palestinians, said he has noticed that real photos and videos of the war tend to have limited reach on Instagram.
“The picture from AI can spread faster in a short time,” he said. Shah says another problem is that he has had graphic images of war removed by Instagram for violating the platform’s policies. He said he’s aware that repeat violations can mean that “users can get blocked,” he said.
Felix Simon, a research fellow at the University of Oxford who studies AI’s impact on public discourse, said the image being created by artificial intelligence fueled its virality far less than other factors.
“The simplicity of the slogan, the symbolism at work, the timing and political context, and the fact that it was shared by celebrities,” said Simon, adding that “the lack of graphic content makes it less likely to get taken down, which helps, too.”
Some commentators criticized the meme for portraying a sanitized version of war that renders human horrors on the ground in Gaza into an easily-shareable AI image.
Both AbKa and Shah reject that idea, saying AI images can be a useful way to grab peoples' attention and make them engage in some way with the war.
Yet there is no agreement among them about who created the viral image that has spurred discussion around the world about the authenticity of online activism and renewed attention on an internet increasingly rife with realistic-looking AI depictions.
When pressed in direct messages on Instagram for a response to AbKa’s contention that her image was copied, Shah blocked an NPR reporter.
NPR researcher Susie Cummings contributed to this report.
