AI News Coverage, Pseudo AI Companies, and more on COVID-19
26:27 Share

Welcome to SkyNet Today's Let's Talk AI podcast, where you can hear from AI researchers about what's actually going on with AI and what is just clickbait headlines. This week, we'll look at an interesting pair of articles on AI journalism and how AI companies are faking their AI-based solutions, and then turn our attention once again to the latest updates in the fight against coronavirus.

You can find the articles we discuss here today and subscribe to our weekly newsletter with similar ones at skynetoday.com.

I am Andrey Kronkov, a third-year PhD student at the Stanford Vision and Learning Lab. I focus mostly on learning algorithms for robotic manipulation. And with me is my co-host. And I'm Sharon, a third-year PhD student in the Machine Learning group working with Andrew Ng. I do research on generative models, improving generalization of neural networks, and applying machine learning to tackling the climate crisis.

And Sharon, you actually did the first interview of this podcast just last week when you talked to Professor Matthew Longren about AI and COVID-19 and medicine. So I'm curious, what did you find particularly interesting in that interview? I found it interesting that he was very optimistic and happy with the tech community and how the tech community is engaging with helping with the crisis quite a bit.

I've definitely heard other opinions from folks perhaps on the other side, thinking that people trying to create their own ventilators may do more harm than good. But I like his optimism and I like hearing his perspective from a doctor's point of view. Yeah, I have seen this counter perspective that people in AI or data science trying to model the disease or writing various things about it

are kind of being armchair experts without any actual medicine knowledge and so are doing more harm than good. And it was nice to hear that, according to him, at least some people in our field are being useful and not just, you know, pretending to know what they're doing. Right, right. It's probably not good to necessarily be doing self-promotion in this time, but I do see people doing that, unfortunately. Yeah.

I guess that's only to be expected. But let's not dwell on COVID-19 too long. We'll have time to talk about it later, as usual these days. But first, we're going to focus more on AI and longer term trends and what's been going on for quite a while. With our first piece being...

titled There's Room to Improve AI News Coverage. And this is from futurity.org. So this piece is an interview with several people who wrote the paper AI in the Headlines, The Portrayal of the Ethical Issues of AI in the Media, which is a paper that was published in AI and Society. And this paper basically looks into the statistics of AI

that have been written in the past several years. They looked at the range from 2013 to 2018, and they kind of analyzed what kind of writing has been done, what are its contents, what is its focus, and they also had some recommendations to how it could be better. Yeah, so the findings were quite interesting. There was a sharp increase in the number of articles published in recent years, specifically with the number of

of articles published between 2013 and 2017

being less than the number of articles published in 2018, that one year alone, suggesting also that the public discussion on AI is fairly new. They also found that no relevant media articles were really found prior to 2013. So the top 10 most common themes that they found were around prejudice, privacy data protection, transparency, job loss to AI, economic impact of AI, AI decision-making,

responsibility or liability, weaponization of AI, AI abuse, embedding ethics in AI, and safety. This is how they clustered the different issues on AI in those media articles. And in addition, they also had some other clusters of recommendations and types of AI in those articles. And I noted that a large portion of recommendations in these media articles had to do with developing ethical guidelines.

which of course has been a pretty large ongoing conversation as AI has been larger. And they also noted with these types of AI coverage that

Most articles don't discuss any specifics. They pretty much broadly discuss the field of AI with sometimes examples of autonomous vehicles or chatbots or surveillance or these sorts of things. In fact, they suggest that journalists should reach out to both AI experts and ethicists to fact check a little bit more. For example, one of the most commonly cited quote unquote ethical frameworks in these articles about AI ethics is Isaac Eberhardt.

Asimov's Three Laws of Robotics, which comes from his famous short stories. It's not a formal ethical framework and it's not very useful in the context of real world discussions about how AI can and should behave ethically. Utilitarianism is also commonly discussed in these news articles about AI ethics, but these are all fairly basic and could be upgraded. We could apply modernization

ethical frameworks that are capable of tackling the, quote, intricacies and complexity of real world applications of artificial moral agents, the article cites. And this part of their findings has to do with something else they note, which is that most of articles had neutral tones. So they were not, let's say, hyperbolic. There was actually not too much hype, not too much fear mongering, but rather they focused on practical and relevant issues having to do with AI.

but they also didn't have a specific accommodations so the conclusion here seems to be

that journalists are doing their job pretty well in terms of being neutral and relevant, but that they could dig deeper by talking to both AI experts and ethicists. And speaking of recommendations, actually, this relates to an article we've had at Skana today as of last year, which is titled AI Coverage Best Practices According to AI Researchers, where we surveyed a bunch of AI researchers and

and basically clustered their recommendations into a set of do and don't guidelines for journalists. And in that article, we really focused on more of a specifics. So in general, people tend to get things right, but with particular wording choices or particular things to highlight and not highlight, there are missteps.

So it looks like this article also makes a point that as you get into specifics, you can try to be a little more careful. So while AI journalism might have some ways to go, it does dig up some really interesting things both in academia and in industry. One recent example comes in a Forbes article titled Artificial or Human Intelligence? Companies Faking AI.

And they essentially talk about how we've seen lots of AI companies claiming or offering to provide AI-based solutions to a huge wide range of business problems.

But because that's pretty hard, some companies are choosing to approach these AI challenges not by scaling back their ambitions, but rather by using humans to do the task that they are otherwise trying to get their AI systems to do. And Forbes calls this pseudo AI or quote unquote faking it. Yeah.

This article discusses some examples, and the general trend is that companies claim to automate some process using software and AI. So things like transcription, appointment scheduling, personal assistant work. But to actually implement these things, instead of having AI, they outsource this work to human workers through, let's say, crowdsource marketplaces. One example being Amazon Mechanical Turk.

And they do so without making it clear that the service is actually being powered by human labor and not by some fancy AI. And maybe you can go into some more examples, Sharon. Yeah, so the article states that CNBC published an article, an additional article, critical of SOFIA, which is the AI robot from Hanson Robotics.

And when that company was approached by CNBC with a list of questions that they wanted to ask, Hansa Robotics instead responded with a list of very specific questions that they were allowed to ask for Sophia. So in a follow-up video, CNBC questions whether the robot is meant to be researched into AI or if it's just a PR stunt.

And even the owner of Hanson robotics has gone on record as saying most media encounters are scripted. I will say here that I would probably take the side of the Hanson robotics folks, uh, as most media encounters are scripted as they are pretty delicate situations for companies that could make or break either their stock or their user base or something. Um, and it's a very quick little glimpse into them. So people try to put on their best performance, uh,

And I think a lot of even just people, human interviews are very, very scripted as well, where they want very specific questions asked of them. But to that point, actually, I think it's true that the scripting is not necessarily the problem. We actually had an article on Sky News Today on Sophia.

And what people often do find problematic is not only that it's scripted, but there is an implication that Sofia the robot can actually handle natural conversation if it were not to be scripted. And the truth is that the technology is not there. So they're presenting it as very advanced, as having capabilities that simply don't exist.

And that is more than just being prepared and polished, which we try to do with demos as well as researchers, right? So we prepare demos to show off our algorithms in controlled settings. But here, there's no real algorithm. They're really playing back audio instead of having the AI they claim to have, or they have a much simpler kind of software system in place.

So that's kind of how there can be a misleading element to something that would usually be maybe not that bad.

Additional examples include the popular calendar scheduling services, X.AI and Clara Labs, which some of you may use. And these were found to both be using humans to schedule appointments and calendar items rather than purely AI solutions. And again, I would say I'm not very surprised by this. I think the only issue that I would

would take with a lot of this is how they are paying those humans, for the most part, paying them at very low wages. And also, I'd be concerned about privacy. I've heard about some privacy issues where actually a human labeler starts to try to find the person that they're labeling data for because they've fallen in love with them or something like that. So I think there are definitely concerns around privacy and fair wage to the humans behind the AI solutions.

But what are your thoughts, Andre, on this?

Yeah, I think that's a good point that in general, it's not necessarily wrong to use humans as part of the solution. In fact, there's a strategic component here where while these companies are developing their AI system, they need data to power it. And there's kind of a chicken and egg problem where if you don't have data, you can't provide the AI service. So there is a clever kind of trick of providing a service and powering it with humans first as you gather data and build out the system.

So that's not necessarily wrong, but if you are being misleading about it, if you are over-promising and, as you said, not disclosing privacy concerns or...

or not paying the human employees fairly, then that can lead to a huge amount of disillusionment with this whole class of companies and investment in AI companies drying up, potentially investment in companies with actually exciting AI solutions not happening, and the public just being wary of AI in general.

So there are, I think, quite a few problems with pseudo-AI if it's not disclosed upfront that it's not necessarily just algorithmic. Yes, and I think that's a really interesting point that the article points out with ethics being important, not just from the viewpoint of providing a solution that matches customer and user expectations, but also about instilling trust in a technology that some are still very much wary of or dubious about.

So it's for the future, also entrepreneurs. It's not just for the companies themselves and it's for the brand of AI completely. At the end of the day, if it's information that others would need, then withholding it is, of course, problematic. But in some cases, you don't necessarily need to disclose all the internals of your solution.

Well, with those two interesting articles about AI trends out of the way, let us get back to what you all already love, which is more discussion of COVID-19. Our first article on that topic here is titled

Even the pandemic doesn't stop Europe's push to regulate AI. And this is from Bloomberg. And it's all about how Europe has for several years been enacting many regulations having to do with AI. So first with GDPR, the General Data Protection Regulation. In 2018, they did it again with communication on AI for Europe, which set forward an ethical and legal set of guidelines that

to ensure AI is developed in a certain way in Europe. And then this year, it was announced in Europe that more laws regulating the development of AI would be rolled out. And this article basically discusses how even with the virus crisis going on right now, that is not being stopped. So why should businesses pay attention to this?

So businesses have seen this before, how the EU has, quote, punched above its weight as an international rulemaker. It's sweeping laws on privacy, which forced both Google and Facebook to change how they collected user data, has set a global standard influencing other countries to follow suit. And the same could hold true right now with AI. Okay.

The U.S. and China are home to the biggest commercial AI companies, including Google, Microsoft, the Beijing-based Baidu, and Shenzhen-based Tencent. But if they want to sell to Europe's customers, Europe's businesses, they might be forced to overhaul their operations yet again.

Yeah, and that all is from the article, which also lays out some of the details on this proposed AI legislation. The good news for the companies is that the EU has mapped out a tiered approach to AI legislation, which means that there are different rules for different levels of risk. So this is good for companies because for less risky systems such as AI-based parking meters,

The EU is proposing less intense measures like voluntary labeling, which are not that hard to comply with. Whereas for high risk applications, that could endanger people's safety or legal status, such as self-driving cars or surgery or things like that.

VAU has outlined mandatory legal requirements, which of course would require more work to work on. So for instance, large companies such as Google could be forced to have their systems tested before being deployed, and they might have to actually

We train our systems and algorithms in Europe with different data sets to guarantee the user's rights are held to regulations. Specifically, DeepMind has been conducting research on the coronavirus. And a lot of that research was using open source data from all over the world.

around the world, but this would be limiting some of those AI models to only using European data and that would significantly limit their capabilities in the current crisis. And that also applies to AI in normal situations, of course.

These checks in Europe on AI could also cause Europe to trail behind some of the pro-innovative quote-unquote economies, and this could also incentivize companies to relocate to other markets with fewer bureaucratic hurdles, of course. But of course,

European side and on the EU side, Margarethe Vestager, the EU's tech czar, says that Brussels is standing firm on this, and it's her plan to generate trust around the technology and how it's deployed so that people are more willing to embrace that innovation. So it's a very different perspective. And of course, on both sides,

Everyone is saying, oh, we want people to adopt AI more, but one is through trust and another is through increased globalization.

Yeah, I think it's safe to say in the US, we do not have this approach so much. The regulation is quite lax and there's really only now conversations about maybe doing something similar to GDPR or something for facial recognition or various things like that. Do you think we should be doing more legislation already, Sharon? Or is it good to let the companies really move fast for now and then instill some more laws?

I think regulation could be very helpful, but I also, my sense is historically the U.S. has not been great at setting standards for regulation. Just by way of how privacy laws or laws in general are set up here, I heard it was kind of like Swiss cheese where you're

you, you have a free market, of course. And then you're like in this very specific example, in case you need to preserve privacy. Um, and the EU takes actually the opposite stance, which is we want blanket privacy and there are exceptions to that privacy. So I do, and perhaps this is just like an AB experiment. You can think of it that way, but, um,

I would say the U.S. is generally much more cautious about rolling out regulation. And I would say that I think the EU has had a more thoughtful discussion on it.

And it is quite useful, of course, that this discussion has been happening. In some cases, like with autonomous driving, it's almost a necessary component before the technology can be rolled out. I don't think any company will be trying to offer a service without some knowledge of what the government expects of it for their service. So the sooner it's made clear, the sooner these services can be offered widely. So let's hope that happens quickly.

So in past weeks, we've covered the fact that AI has many useful applications in the fight against coronavirus. A new article from Bloomberg Businessweek on this topic made an interesting point that we'll mention in the context of concerns about AI use. So the title is called The Virus Gives AI a Chance to Prove It Can Be a Force for Good.

Which states, quote, the pandemic is opening up a massive opportunity for the tech industry while it shines a light on calls for more scrutiny of AI innovations being developed faster than regulators are able to devise rules to protect citizens' rights.

Yeah, so we discussed that exactly last week already. You can go to our previous episodes for more details on this. We're just going to move on to our next article, which is from ZDNNet and is titled AI Runs Smack Up Against a Big Data Problem in COVID-19 Diagnosis. And this one goes a bit more into details on a specific application of AI for COVID-19.

So this specific application has to do with diagnosis and in particular using computer vision to diagnose if someone has COVID-19 from images that are derived from x-rays or CT scans of their lungs.

So in China, there have been articles circulated describing amazing successes in a number of AI diagnosis projects of the like. And these include from various companies, including Chinese software maker, Infravision, Chinese insurance firm Ping An's healthcare division, Chinese search giant Alibaba, and Chinese tech startups Deepwise Technology and iFlyTech.

And the article states that the reality is slightly less exhilarating as despite some success, numerous efforts actually face quite a few challenges, especially with the novel disease. Especially with a novel disease, the presence of distinguishing features isn't always conclusive. So when we think we see evidence for a positive, it may not actually be conclusive. So those

Labels are essentially noisy. The article also talks about how a team of over 30 researchers were able to build a deep learning system to read CT scans, and they deployed this to 16 hospitals across China, including one in Wuhan, the epicenter of the COVID-19 crisis, and that achieved a rate of 1,300 screenings per day.

And so that initial success is fantastic, but it's coming up against the reality that it can be hard to move forward in other countries with this technology. Yes, and the article goes on to detail some of the difficulties associated with developing this technology. One of them being actually getting the data to train the algorithms. So obviously you need to have

images of x-rays or CT scans annotated by actual physicians with their conclusions from the image. And that is very time consuming and hard to procure. And AI researchers are not actually able to do it themselves.

So it goes into quite a bit of detail on how this is a difficulty in this area. And of course, physicians are already super busy due to responding to the crisis. So that makes it even harder right now. Having worked in this area for a bit, I can also comment that

physician annotations do not necessarily agree with each other, meaning that one doctor may not see one x-ray as positive, meaning that one doctor may see an x-ray or CT scan as positive, whereas another doctor might see it as negative. And also the definition of positive and negative here are very, very different, or the definition of positive and negative here are not very certain, even among physicians.

medical staff and it's hard to define and it's constantly evolving, which makes this particularly challenging when we do train models on it. And that points to an additional challenge, which is that you actually want to integrate this AI tool with a radiologist or a medician's workflow, which in itself is non-trivial. So ideally, the AI system wouldn't just spit out an answer. It would provide an initial screening or somehow assist the

the actual medical professional in making the call, but that then requires the AI researchers to actually talk to medical practitioners to know what they need.

In addition to the lack of data, in a paper last month, scholars from the World Health Organization, the United Nations Global Pulse, and Montreal's Mila Institute for AI surveyed the landscape of AI applications from diagnosis to potential cures, including x-ray and CT scan software. And their conclusion was that while AI can help in the fight against COVID-19 in

many, many different ways. Many of the systems being developed lack operational maturity, meaning that there is no central database for them to host different things. There's no encryption for privacy on that data. And this article builds on that and basically makes the point that

This is now a time of challenge for AI to kind of try and grow up and deal with this very critical context where getting data is not easy and that it's not easy, but that it will help the field and the technology mature and ultimately make AI useful in the context and future crises, hopefully.

And that will go ahead and finish up this week's episode. Thank you so much for listening to our episode of Skynet Today's Let's Talk AI podcast. You can find the articles we discussed here today and subscribe to our weekly newsletter with similar ones at skynettoday.com. Subscribe to us. Okay. Subscribe to us wherever you get your podcasts and don't forget to leave us a rating if you like the show. Be sure to tune in next week.