Trustworthy AI?

Can We Build Trustworthy AI?  in Gizmodo

AI isn't transparent, so we should all be preparing for a world where AI is not trustworthy, write two Harvard researchers.

By  Nathan Sanders and Bruce Schneier,    Published  May 4, 2023

We will all soon get into the habit of using AI tools for help with everyday problems and tasks. We should get in the habit of questioning the motives, incentives, and capabilities behind them, too.

Imagine you’re using an AI chatbot to plan a vacation. Did it suggest a particular resort because it knows your preferences, or because the company is getting a kickback from the hotel chain? Later, when you’re using another AI chatbot to learn about a complex economic issue, is the chatbot reflecting your politics or the politics of the company that trained it?

Related Content

0:09 / 5:02

Alexa Will Read to Your Kids, but There's a Catch

Amazon Alexa Told a 10-Year-Old Girl to Play With a Live Wall Outlet

For AI to truly be our assistant, it needs to be trustworthy. For it to be trustworthy, it must be under our control; it can’t be working behind the scenes for some tech monopoly. This means, at a minimum, the technology needs to be transparent. And we all need to understand how it works, at least a little bit.

Amid the myriad warnings about creepy risks to well-being, threats to democracy, and even existential doom that have accompanied stunning recent developments in artificial intelligence (AI)—and large language models (LLMs) like ChatGPT and GPT-4—one optimistic vision is abundantly clear: this technology is useful. It can help you find information, express your thoughts, correct errors in your writing, and much more. If we can navigate the pitfalls, its assistive benefit to humanity could be epoch-defining. But we’re not there yet.

Let’s pause for a moment and imagine the possibilities of a trusted AI assistant. It could write the first draft of anything: e-mails, reports, essays, even wedding vows. You would have to give it background information and edit its output, of course, but that draft would be written by a model trained on your personal beliefs, knowledge, and style. It could act as your tutor, answering questions interactively on topics you want to learn about—in the manner that suits you best and taking into account what you already know. It could assist you in planning, organizing, and communicating: again, based on your personal preferences. It could advocate on your behalf with third parties: either other humans or other bots. And it could moderate conversations on social media for you, flagging misinformation, removing hate or trolling, translating for speakers of different languages, and keeping discussions on topic; or even mediate conversations in physical spaces, interacting through speech recognition and synthesis capabilities.  ...'

How Trustless is Bitcoin?

Full Article

How 'Trustless' Is Bitcoin, Really?

The New York Times. Siobhan Roberts, June 6, 2022

Rice University's Alyssa Blackburn and colleagues have dissected the anonymity of bitcoin, reporting in a paper that "information leakage erodes the once-impenetrable blocks, carving out a new landscape of socioeconomic data." The researchers aggregated multiple leakages and bitcoin addresses, determining 64 key agents mined most of the existing bitcoin in the first two years since the cryptocurrency's launch. Blackburn devised hacks for this period, and tapped human lapses like insecure user behavior, operational features innate to bitcoin's software, and methods for connecting pseudonymous addresses. She said very few people serve as network arbiters, "which is not the ethos of decentralized trustless crypto." Blackburn also noted that the concentration of resources undercut the network's security, with a miner's computing resources found to be commensurate to their mining income.

Full Article

A Cyber Security Social Contract

The current state of the affairs', and related threats are pointing to a need for  this.

The Cyber Social Contract  in Foreign Affairs

How to Rebuild Trust in a Digital World

By Chris Inglis and Harry Krejsa, February 21, 2022

In the spring of 2021, a Russia-based cybercrime group launched a ransomware attack against the largest fuel pipeline in the United States. According to the cybersecurity firm Mandiant, the subsequent shutdown and gas shortage across the East Coast likely originated from a single compromised password. That an individual misstep might disrupt critical services for millions illustrates just how vulnerable the United States’ digital ecosystem is in the twenty-first century.

Although most participants in the cyber-ecosystem are aware of these growing risks, the responsibility for mitigating systemic hazards is poorly distributed. Cyber-professionals and policymakers are too often motivated more by a fear of risk than by an aspiration to realize cyberspace’s full potential. Exacerbating this dynamic is a decades-old tendency among the large and sophisticated actors who design, construct, and operate digital systems to devolve the cost and difficulty of risk mitigation onto users who often lack the resources and expertise to address them.

Too often, this state of affairs produces digital ecosystems where private information is easily accessible, predatory technology is inexpensive, and momentary lapses in vigilance can snowball into a continent-wide catastrophe. Although individually oriented tools like multifactor authentication and password managers are critical to solving elements of this problem, they are inadequate on their own. A durable solution must involve moving away from the tendency to charge isolated individuals, small businesses, and local governments with shouldering absurd levels of risk. Those more capable of carrying the load—such as governments and large firms—must take on some of the burden, and collective, collaborative defense needs to replace atomized and divided efforts. Until then, the problem will always look like someone else’s to solve.

The United States needs a new social contract for the digital age—one that meaningfully alters the relationship between public and private sectors and proposes a new set of obligations for each. Such a shift is momentous but not without precedent. From the Pure Food and Drug Act of 1906 to the Clean Air Act of 1963 and the public-private revolution in airline safety in the 1990s, the United States has made important adjustments following profound changes in the economy and technology.  ...

See also comments in Schneier:  https://www.schneier.com/blog/archives/2022/02/a-new-cybersecurity-social-contract.html 

Trusting Blockchain

This article was pointed to me by the immediately preceding one,  See also the comments at the link for further debate. 

There's No Good Reason to Trust Blockchain Technology

Bruce Schneier, in Wired, February 6, 2019

In his 2008 white paper that first proposed bitcoin, the anonymous Satoshi Nakamoto concluded with: “We have proposed a system for electronic transactions without relying on trust.” He was referring to blockchain, the system behind bitcoin cryptocurrency. The circumvention of trust is a great promise, but it’s just not true. Yes, bitcoin eliminates certain trusted intermediaries that are inherent in other payment systems like credit cards. But you still have to trust bitcoin—and everything about it.

Much has been written about blockchains and how they displace, reshape, or eliminate trust. But when you analyze both blockchain and trust, you quickly realize that there is much more hype than value. Blockchain solutions are often much worse than what they replace.  .... ' 

Should We Trust Computers?

 Good talk  given by Prof Martyn Thomas  CBE in the Gresham College lecture series  published Oct 29 2015.   https://www.youtube.com/watch?v=8SZfjvlbpMw

Trustworthy AI Blog

IBM Writes on Trustworthy AI in a new Blog:

Financial services: Trustworthy AI’s promise and payoff

IBM, Forrester, UBS, Regions Bank, and State Bank of India show how to build AI responsibly

By Jennifer Clemente 

Data: It’s becoming richer, cheaper, and increasingly open in the post-pandemic world of broad digitalization. And for financial services organizations, this windfall is opening pathways to improve business performance with end-to-end intelligent automation and hyper-personalized services to satisfy customers’ expectations.

As data gets cheaper, trust grows in value

For banks, insurance companies and other financial institutions on a quest for reinvention, the pressure to play catch up, or leapfrog means fast-tracking AI deployments while balancing governance, risk and compliance needs.

According to a January 2020 Forrester Consulting study commissioned by IBM, Overcome Obstacles to get to AI at scale, 40 percent of the participants report data governance issues are a serious concern. Additionally, 58 percent of its participants indicated that data quality issues are the number one challenge in their organization.

Now the good news.

A trusted architecture based on data and AI solutions generates higher business value, according to Brandon Purcell, Vice President and Principal Analyst at Forrester who spoke at a recent Data and AI Virtual Forum keynote, Trustworthy AI: Forging the future of banking, insurance and financial markets.  .... ' 

Communities and Trust in Pandemic Context

 Interesting research here out of Kellogg Northwestern.  In the process of reading. 

Trust Usually Helps Communities Thrive. During a Pandemic, Not So Much.

Places with high levels of trust are worse at social distancing.

BASED ON THE RESEARCH OF  Georgy Egorov, Ruben Enikolopov ,Maria Petrova

In general, high levels of trust are considered good for society. And one would assume this to be particularly true during a pandemic: if we trust that our neighbors are following isolation and social-distancing guidelines, wouldn’t we follow them as well?  ... 

Border Gateway Protocol

 New to me, a useful description here.

Fixing the Internet  By Keith Kirkpatrick

Video description:  https://youtu.be/A1KXPpqlNZ4  

Communications of the ACM, August 2021, Vol. 64 No. 8, Pages 16-17  10.1145/3469287

Few people pay much attention to how the electrical grid works until there is an outage. The same is often true for the Internet.

Yet unlike the electrical grid, where direct attacks are infrequent, vulnerabilities and security issues with the Internet's routing protocol have led to numerous, frequent malicious attacks that have resulted in widespread service outages, intercepted and stolen personal data, and the use of seemingly legitimate Web sites to launch massive spam campaigns.

The Internet is an interconnected global network of autonomous systems or network operators, like Internet service providers (ISPs), corporate networks, content delivery networks (such as Hulu or Netflix), and cloud computing companies such as Google and Microsoft Cloud. The Border Gateway Protocol (BGP) is used to ensure data can be directed between networks along the most efficient path, similar to how a GPS navigation system maintains a database of street addresses and can assess distance and congestion when selecting the optimal route to a destination.

Each autonomous system connected to the Internet has an Internet Protocol (IP) address, which is its network interface, and provides the location of the host within the network; this allows other networks to establish a path to that host. BGP routers managed by an ISP control the flow of data packets containing content between networks, and maintains a standard routing table used to direct packets in transit. BGP makes routing decisions based on paths, rules, or network policies configured by each network's administrator.

BGP was first described in a document assembled by the Internet Society's Network Working Group in June 1989 and was first put into use in 1994. BGP is extremely scalable, allowing tens of thousands of networks around the world to be connected together, and if a router or path becomes unavailable, it can quickly adapt to send packets through another reconnection. However, because the protocol was designed and still operates on a trust model that accepts that any information exchanged by networks is always valid, it remains susceptible to issues such as information exchange failures due to improperly formatted or incorrect data. BGP can also be at the mercy of routers too slow to respond to updates, or that run out of memory or storage, situations that can cause network timeouts, bad routing requests, and processing problems.   ....   ' 

FBI Running Encrypted Phones

Bruce Schneier posted about the FBI effort to create an encrypted phone App.  And makes interesting points about trust and security.    Just read, worth considering

FBI/AFP-Run Encrypted Phone     Bruce Schneier

For three years, the Federal Bureau of Investigation and the Australian Federal Police owned and operated a commercial encrypted phone app, called AN0M, that was used by organized crime around the world. Of course, the police were able to read everything — I don’t even know if this qualifies as a backdoor. This week, the world’s police organizations announced 800 arrests based on text messages sent over the app. We’ve seen law enforcement take over encrypted apps before: for example, EncroChat. This operation, code-named Trojan Shield, is the first time law enforcement managed an app from the beginning.

If there is any moral to this, it’s one that all of my blog readers should already know: trust is essential to security. And the number of people you need to trust is larger than you might originally think. For an app to be secure, you need to trust the hardware, the operating system, the software, the update mechanism, the login mechanism, and on and on and on. If one of those is untrustworthy, the whole system is insecure. ...  '

It’s the same reason blockchain-based currencies are so insecure, even if the cryptography is sound.  ...'

Measuring Trust in AI

Accurate measures in context would be very useful.

NIST Wants to Measure Trust in AI   By Wired, June 22, 2021

The National Institutes of Standards and Technology wants to quantify user trust in artificial intelligence.

"Trust and Artificial Intelligence," by Brian Stanton at NIST's Information Technology Laboratory and Theodore Jensen at the University of Connecticut, aims to help businesses and developers who deploy AI systems make informed decisions and identify areas where people don't trust AI.

Without trust, Stanton says, adoption of AI will slow or halt.

NIST views the initiative as an extension of its more traditional work establishing trust in measurement systems. Public comment is being accepted until July 30.

From Wired  

Criticism of Blockchains and Trust

Here's a 2019 paper/blog post by Bruce Schneier that I read then.    Now revisiting the idea of trust and blockchain.   Article is usefully critical of blockchain in general, but some good thoughts about elements of trust.   Very critical, especially regards the aspects of trust in monetary systems.    There are many, many links, and many. many comments.     Go through the link below for access to comments and links. Considerable piece.

Blockchain and Trust  By Bruce Schneier

In his 2008 white paper that first proposed bitcoin, the anonymous Satoshi Nakamoto concluded with: “We have proposed a system for electronic transactions without relying on trust.” He was referring to blockchain, the system behind bitcoin cryptocurrency. The circumvention of trust is a great promise, but it’s just not true. Yes, bitcoin eliminates certain trusted intermediaries that are inherent in other payment systems like credit cards. But you still have to trust bitcoin — and everything about it.

Much has been written about blockchains and how they displace, reshape, or eliminate trust. But when you analyze both blockchain and trust, you quickly realize that there is much more hype than value. Blockchain solutions are often much worse than what they replace.

First, a caveat. By blockchain, I mean something very specific: the data structures and protocols that make up a public blockchain. These have three essential elements. The first is a distributed (as in multiple copies) but centralized (as in there’s only one) ledger, which is a way of recording what happened and in what order. This ledger is public, meaning that anyone can read it, and immutable, meaning that no one can change what happened in the past.

The second element is the consensus algorithm, which is a way to ensure all the copies of the ledger are the same. This is generally called mining; a critical part of the system is that anyone can participate. It is also distributed, meaning that you don’t have to trust any particular node in the consensus network. It can also be extremely expensive, both in data storage and in the energy required to maintain it. Bitcoin has the most expensive consensus algorithm the world has ever seen, by far.

Finally, the third element is the currency. This is some sort of digital token that has value and is publicly traded. Currency is a necessary element of a blockchain to align the incentives of everyone involved. Transactions involving these tokens are stored on the ledger.

Private blockchains are completely uninteresting. (By this, I mean systems that use the blockchain data structure but don’t have the above three elements.) In general, they have some external limitation on who can interact with the blockchain and its features. These are not anything new; they’re distributed append-only data structures with a list of individuals authorized to add to it. Consensus protocols have been studied in distributed systems for more than 60 years. Append-only data structures have been similarly well covered. They’re blockchains in name only, and — as far as I can tell — the only reason to operate one is to ride on the blockchain hype.  ... 

What should a Robot do when it Cannot Trust the Model it was Trained on?

Also a thing we expect of useful 'intelligence'... knowing its limitations.  Or do we?  How is this different?  

What should a robot do when it cannot trust the model it was trained on?

Helping Robots Learn What They Can and Can't Do in New Situations

The Michigan Engineer News Center, Dan Newman, May 19, 2021

University of Michigan researchers have developed a method of helping robots to predict when the model on which they were trained is unreliable, and to learn from interacting with the environment. Their approach involved creating a simple model of a rope's dynamics while moving it around an open space, adding obstacles, creating a classifier that learned when the model was reliable without learning how the rope interacted with the objects, and including recovery steps for when the classifier determined the model was unreliable. The researchers found their approach was successful 84% of the time, versus 18% for a full dynamics model, which aims to incorporate all possible scenarios. The approach also was successful in two real-world settings that involved grabbing a phone charging cable, and manipulating hoses and straps under a car hood. Michigan's Dmitry Berenson said, "This method can allow robots to generalize their knowledge to new situations that they have never encountered before."  ... ' 

Accenture: Get a Secure, Trustworthy Internet

Good, considerable piece by Accenture, pointed to below.   Who we worked with in the past.  Have been of late become involved in related efforts and find some their points key.   As has been reported here.   How do we secure the digital internet, and make people trust it for all tasks,  large and small, personal and commercial?   We are slipping here and need to get a better handle on the risks,  included losing our hold.   Else what?  

Securing the foundations

The benefits of a secure, trustworthy Internet economy are clear.

CEOs have an opportunity to drive meaningful change today and develop a foundation of trust for tomorrow’s digital economy. Unfortunately, just one attack is all it takes to damage an organization.

The actions of CEOs—driving above ground and influencing below ground—matter. By joining forces with other CEOs, public sector leaders and regulators, they can develop much-needed guidelines and oversight mechanisms. By protecting their own organization and extending protection through its value chain, they will safeguard the business ecosystem. By embracing and developing technologies that can advance their businesses and enhance digital safety, CEO engagement can drive a trust turnaround for the Internet and secure the future of the digital economy. ... " 

From Blockchain to Contracts

Have been looking at the use of 'smart contracts' in a broader way.  Is this one direction?   Incorporating trust.

Investing in Aleo  by Katie Haun and Ali Yahya  in Andreessen Horowitz

From the beginning, our core thesis has been that the best way to think of a modern blockchain is as a new class of computer that has the ability to run a special kind of program. These programs are sometimes called smart contracts, and they’re different from ordinary programs in that they have a life of their own. They are independent, and once written, they obediently execute themselves subject to nobody’s authority. Because of this property, smart contracts are uniquely capable of earning trust. 

But smart contracts today have two big limitations: (1) they are fully transparent by design and therefore don’t allow for privacy (2) they don’t scale to millions (let alone billions) of users. These limitations exist because trust requires verification. Transactions on a blockchain need to be transparent so that everyone can verify that they are correct. And, they tend not to scale because it takes time and energy for all computers on the network to perform that verification.

But research in a cutting edge area of cryptography called zero-knowledge proofs promises to unlock an elegant solution to the privacy and scalability problems. We spent a great deal of time looking at various approaches and teams working on this. Aleo’s solution is both elegant and pragmatic.  ... '

On Recommendation Systems

 Anther well considered piece, on recommendation systems, link to original post includes more useful references.  Note the inclusion of elements of 'trust'.  Intro below. 

The Increasing Influence of Recommendation Systems in Our Everyday Lives by Irving Wladlensky-Berger in his blog ... 

A few years ago, I attended a seminar by University of Toronto professor Avi Goldfarb on the economic value of AI. Goldfarb explained that the best way to assess the impact of a new radical technology is to look at how the technology reduces the cost of a widely used function. Computers, for example, are powerful calculators whose cost of arithmetic and other digital operations have dramatically decreased over the past several decades. As a result, we’ve learned to define all kinds of tasks in terms of digital operations, e.g., financial transactions, inventory management, word processing, photography. Similarly, the Internet and World Wide Web have drastically reduced the cost of communications and of access to all kinds of information, - including numbers, text, pictures, music and videos.

Viewed through this lens, the data and AI revolution can be viewed as reducing the cost of predictions. Predictions mean anticipating what is likely to happen in the future. Over the past decade, increasingly powerful computers, advanced machine learning algorithms, and the explosive growth of big data have enabled us to extract insights from the data and turn them into valuable predictions. As was previously the case with digital operations, communications and access to information, - we’re now able to reframe all kinds of applications as prediction problems. A major such family of applications are recommendation engines or recommender systems, which Wikipedia defines as “a subclass of information filtering system that seeks to predict the ‘rating’ or ‘preference’ a user would give to an item.”  ... '

Global Variances in Digital Trust

 Digital Trust, supposedly accurately measured.  Had not seen this before, probably useful if accurate.   The components of the measure also seem  they would be hard to generally measure.   And likely have   considerable variance over time, depending on the news.  Below a summary of a long article in the   HBR.   

How Digital Trust Varies Around the World  by Bhaskar Chakravorti, Ajay Bhalla, and Ravi Shankar Chaturvedi     February 25, 2021

Summary.   

As economies around the world digitalize rapidly in response to the pandemic, one component that can sometimes get left behind is user trust. What does it take to build out a digital ecosystem that users will feel comfortable actually using? To answer this question, the authors explored four components of digital trust: the security of an economy’s digital environment; the quality of the digital user experience; the extent to which users report trust in their digital environment; and the extent to which users actually use the digital tools available to them. They then used almost 200 indicators to rank 42 global economies on their performance in each of these four metrics, finding a number of interesting trends around how different economies have developed mechanisms for engendering trust, as well as how different types of trust do — or don’t — correspond to other digital development metrics.   ...' 

Robots Encourage Risk-Taking Behavior in Humans

 Fascinating premise, but highly context dependent,  would not go into a project with this idea strongly held.  Has some relationship to the 'Media Equation' concept, where trust where we trust robots more because they apparently have no hidden motives.    We used that method to drive us to assistant models. 

'The Robot Made Me Do It': Robots Encourage Risk-Taking Behavior in Humans By University of Southampton

Robots can encourage humans to take greater risks in a simulated gambling scenario than they would if the robot is silent, research shows. Increasing understanding of whether robots can affect risk-taking could have clear ethical, practical, and policy implications, according to "The Robot Made Me Do It: Human-Robot Interaction and Risk-Taking Behavior," published in Cyberpsychology, Behavior, and Social Networking.

The research involved 180 undergraduate students taking the Balloon Analogue Risk Task (BART), a computer assessment that asks participants to press the spacebar on a keyboard to inflate a balloon displayed on the screen. With each press of the spacebar, the balloon inflates slightly, and 1 penny is added to the player's "temporary money bank." The balloons can explode randomly, meaning the player loses any money they have won for that balloon and they have the option to "cash-in" before this happens and move on to the next balloon.

One-third of the participants took the test in a room on their own, one third took the test alongside a robot that only provided instructions but was silent the rest of the time, and the third group took the test with the robot providing instruction as well as speaking statements such as, "Why did you stop pumping?"

The results showed that the group who were encouraged by the robot took more risks, blowing up their balloons significantly more frequently than those in the other groups did. They also earned more money overall.

"On the one hand, our results might raise alarms about the prospect of robots causing harm by increasing risky behavior," says Yaniv Hanoch, associate professor in risk management at the University of Southampton who led the study. "On the other hand, our data points to the possibility of using robots, and AI, in preventive programs such as anti-smoking campaigns in schools, and with hard to reach populations, such as addicts."

From University of Southampton

Smarter Models Include Uncertainty, Risk

In the earlier days of AI we always included uncertainty and risk models in parallel.  Even if risk was apparently minimal.    This seems to be much less done today.    Except if you use Bayesian methods that directly include uncertainty.   Are we just not willing to accept risk in stronger context? In the below a claim they are doing it, will take a further look.   'Smart' should always mean understanding uncertainty and risk.

Smarter Models, Smarter Choices    By University of Delaware 

Researchers at the University of Delaware and the University of Massachusetts-Amherst have published details of a new approach to artificial intelligence that builds uncertainty, error, physical laws, expert knowledge, and missing data into its calculations and leads ultimately to much more trustworthy models.

Researchers at the universities of Delaware (UD) and Massachusetts-Amherst have developed a high-confidence approach to artificial intelligence-based models that incorporates uncertainty, error, physical laws, expert knowledge, and missing data into its calculations.

The model itself identifies data required to reduce errors, enabling a higher level of theory for generating more accurate data, further shrinking error boundaries on predictions and the area to explore.

UD's Joshua Lansford said, "Uncertainty is accounted for in the design of our model. Now it is no longer a deterministic model. It is a probabilistic one."

From University of Delaware

People Too Trusting of Virtual Assistants

Another look at the general idea of creating a 'personality' of assistants.   Anthropomorphic features like names, speaking style, voice, gender,  etc.  can influence how we perceive trust, truth, value, uses.     Of course just because things look more like humans does not create required trust.  Again, many issues of needed context are important here. Combining human and machine capabilities is now often being used.

People Too Trusting of Virtual Assistants
University of Waterloo News

Researchers at the University of Waterloo in Canada have found that people tend to share increasingly more with online agents because of their tendency to assign them personalities and physical features like age, facial expressions, and hairstyles. The researchers asked 10 men and 10 women to interact with three conversational agents—Alexa, Google Assistant, and Siri. The team then interviewed each participant to determine their perception of the agents' personalities and what they would look like, before asking each person to create an avatar for each agent. The researchers found Siri was most frequently described as disingenuous and cunning, while Alexa was perceived as genuine and caring. Said University of Waterloo researcher Anastasia Kuzminykh, "How an agent is perceived impacts how it's accepted and how people interact with it; how much people trust it, how much people talk to it, and the way people talk to it."

On the Trust of AI

Over time this will change, but mere statistical statements of performance are still hard to win with.  Augmented a doctor's reasoning capability is still a more powerful argument.   Note the argument of 'uniqueness' of their case mentioned here.   The doctor is still there to address unique cases.

AI Can Outperform Doctors. So Why Don’t Patients Trust It?   By Chiara Longoni,  Carey K. Morewedge in HBR

Our recent research indicates that patients are reluctant to use health care provided by medical artificial intelligence even when it outperforms human doctors. Why? Because patients believe that their medical needs are unique and cannot be adequately addressed by algorithms. To realize the many advantages and cost savings that medical AI promises, care providers must find ways to overcome these misgivings.

Medical artificial intelligence (AI) can perform with expert-level accuracy and deliver cost-effective care at scale. IBM’s Watson diagnoses heart disease better than cardiologists do. Chatbots dispense medical advice for the United Kingdom’s National Health Service in lieu of nurses. Smartphone apps now detect skin cancer with expert accuracy. Algorithms identify eye diseases just as well as specialized physicians. Some forecast that medical AI will pervade 90% of hospitals and replace as much as 80% of what doctors currently do. But for that to come about, the health care system will have to overcome patients’ distrust of AI. .... "