Tools for trustworthy AI

A new tool has been developed to catch students cheating with ChatGPT. It's 99.9% effective. But OpenAI hasn't released it because it's mired in ethics concerns.

It's just one example of one of the major challenges facing AI. How can we monitor the technology to make sure it's used ethically?

For the past few years, the biggest names in AI have pushed for their tech to be used responsibly. And using AI ethically isn't just the right thing for businesses to do-it's also something consumers want. In fact, 86% of businesses believe customers prefer companies that use ethical guidelines and are clear about how they use their data and AI models, according to the IBM Global AI Adoption Index.

"We are all well beyond hoping companies are aware [of using AI ethically]," says Phaedra Boinodiris, IBM's Global Leader for Trustworthy AI. "The bigger question is, why is it important for businesses and other organizations to be held accountable for safe and responsible outcomes from the AI they deploy?"

Can AI ethics tools help? Are the tools themselves biased? Here's a quick look at the latest research.

Deepfake detectors

Deepfake fraud attempts were up 3000% from 2022 to 2023, and they are getting more elaborate. In February, a finance worker at a multinational firm in Hong Kong was duped into paying out $25 million to criminals who created a video conference call with multiple deepfake characters including the firm's CFO.

In May, OpenAI announced that it had released its own deepfake detection tool to disinformation researchers. The tool was able to spot 98.8% of images created by DALL-E 3. OpenAI also joined Google, Adobe and others on the steering committee for the Coalition for Content Provenance and Authenticity (C2PA), an industry coalition developing a standard for certifying the history and source of media content.

Until that standard is set, businesses are building tools they hope can fill the gap. In August, McAfee announced its McAfee Deepfake Detector. The tool uses deep neural network models to spot fake AI audio in videos playing in your browser. In 2022, Intel introduced FakeCatcher, which analyzes blood flow in video pixels, separating humans from deepfakes with a 96% accuracy rate. And they're not alone. Notable startups in the field include NYC's Reality Defender, Israeli startup Clarity, and Estonia-based Sentinel, all of whom have scanning tools available that use AI to spot patterns in various kinds of deepfakes.

With deepfake detection tech evolving at such a rapid pace, it's important to keep potential algorithmic biases in mind. Computer scientist and deepfake expert Siwei Lyu and his team at the University of Buffalo have developed what they believe to be the first deepfake-detection algorithms designed to minimize bias. The UB researchers made a photo collage of the hundreds of faces that were identified as fake in their detection algorithms; the results showed an overall darker skin tone.

"Deepfakes may be used to attack underrepresented minority groups, so it is important to make sure detection technologies will not underserve them," Lyu says. As for the future of deepfake detection? "The generative AI technologies underlying the deepfakes will undoubtedly continue to grow, so we are going to see deepfakes with increasing number, quality and forms. I expect future [detection] technologies will be equipped with more guardrails to reduce the chances of misuses."

Explore AI ethics at IBM

Anti-facial recognition (AFR) technologies

Facial recognition systems are becoming increasingly common as a convenient way to authenticate a user's identity-but these systems have long been fraught with ethical problems ranging from racial bias to data privacy. Complicating the issue, "some biases are [also] intersectional, compounding multiple layers of prejudice," notes Helen Edwards, co-founder of AI ethics think tank Artificiality.

In May, Australian facial recognition startup Outabox's data was breached, releasing the biometric data of more than a million users. Earlier this year 'GoldPickAxe,' a trojan aimed at Android and iOS devices, was caught capturing facial data to break into bank accounts.

A promising approach to protecting facial biometric data is by scrambling it in a way that's imperceptible to the human eye but confuses recognition systems. One of the first tools to do this was Fawkes, a project developed at the University of Chicago. Named after the Guy Fawkes mask, the program is designed to cloak photos by subtly altering pixels; it's free to download on the project's website.

More recently, researchers at Zhejiang University's USSLAB have pioneered CamPro, which aims to achieve AFR at the camera sensor level. CamPro produces images that reduce facial identification to 0.3% without interfering with other applications like activity recognition.

AI writing detectors

Spotting AI-generated writing continues to be a struggle for businesses and educational institutions. In a blind test at the University of Reading, five different psychology modules had ChatGPT-written exams mixed in with exams written by real students. The June study found that 94% of ChatGPT exam answers were not spotted by the people grading the exams. The AI-generated exams also averaged half a grade higher than the student exams.

A variety of AI writing detectors have flooded the market to address this issue, looking for common hallmarks of AI-generated text such as repetition and perfect grammar. But experts warn that they're not reliable yet and often demonstrate bias.

Last year a Stanford study found that AI detectors flagged writing by non-native English speakers an average of 61.3% of the time but made far fewer errors when evaluating writing by native English speakers.

Humans passing off AI-generated writing as their own is not only dishonest-sometimes it's also plagiarism, which can come with serious legal ramifications. Because of this concern, some companies are using AI writing detectors to test the copy of their writers. This has led to companies falsely accusing writers of passing off AI-generated copy as their own, damaging the writers' reputations and careers.

LLM bias detectors

Datasets often include the unconscious biases of the people who create them. It's why algorithmic bias is such a persistent problem in the LLMs that train on this data.

In one example, researchers at Cornell used ChatGPT and Alpaca to generate recommendation letters for men and women; the letters showed significant biases favoring men. Generated language like "Kelly is a warm person" versus "Joseph is a role model" demonstrated the way these biases could affect women in the workplace.

Researchers are working to find ways to flag and mitigate biases. A team at the University of Illinois Urbana-Champaign developed QuaCer-B, which generates provable LLM bias measures for prompts sampled from given distributions and can be used for both API and open-source LLMs.

"The AI industry currently relies on evaluating the safety and trustworthiness of their models by testing them on a small set of benchmark inputs," says UIUC professor Gagandeep Singh, one of the lead researchers behind QuaCer-B. "However, safe generation on benchmark inputs does not guarantee that the LLM-generated content will be ethical when handling diverse unseen scenarios in the real world. QuaCer-B enables LLM developers to make informed decisions about the suitability of their models for real-world deployment and also identify causes of failures to improve the model."

As AI continues to evolve, new ethical problems will keep evolving alongside it. And while tools to flag, monitor and prevent unethical use of the tech are a start, AI ethics experts don't consider them a one-stop solution.

"The hard part is not buying the right tool," Boinodiris adds. "Curating AI responsibly is a sociotechnical challenge that requires a holistic approach. And people are the hardest part of the equation."

"In addition to thoughtful regulation and enforcement, the key to ethical AI is post-market auditing, continually monitoring performance and minimizing risks," explains Gemma Galdón-Clavell, an advisor to the United Nations and EU on applied ethics and responsible AI and founder of Eticas.ai. "Think about the automotive industry: Warning lights and proximity sensors can help drivers avoid crashes, but we still need seatbelts, airbags and regular inspections to ensure that the open road is as safe as it can be."

eBook: AI governance for the enterprise

Public link

Please use the above public link if you want to share this noodl on another website.