Dr. Tedros, I Can’t Let You Do That: The WHO’s Strange AI Chat Bot

A recent tweet by the inimitable Charles Gardner alerted me to the World Health Organization’s AI bot, SARAH. I was pretty surprised to learn about it, chiefly because I didn’t know the WHO was so devoid of cold, calculating, and charisma-free representatives that it felt the urgent need to create even more.

But if 2001: A Space Odyssey taught me anything about AI, it’s that you can’t always trust technology to act in the ways you want. So, I’m going to try to corrupt SARAH, take her off the reservation, and get her to betray her parents —something I haven’t tried in person since I was a young, less balding man.

What is S.A.R.A.H?

OK, so first, here is a bit of background: SARAH, short for Smart AI Resource Assistant for Health, is a “digital health assistant” launched in April 2024. The current iteration evolved from a project called Florence, which the WHO initially introduced to counter misinformation about COVID-19 and help people stop smoking.

The current version of SARAH speaks eight languages and is powered by generative AI. Her scope has also increased considerably, and she now offers advice on healthy eating, physical and mental health, and vaping.

However, SARAH is not just text bot. She also has a voice and a digital image, which are provided by Soul Machines, a US company that creates avatars that don’t quite escape the uncanny valley. Though no longer listed on their website, a look through Wayback Machine reveals that Soul Machines made celebrity avatars for stars like Marilyn Monroe, Francis Ngannou, and Jack Nicklaus.

Sadly, SARAH hasn’t received the best reception. Shortly after launch, Politico magazine ran the rule over the AI bot and found SARAH “wildly inconsistent, often giving contradictory answers to the same queries.” In the same article, they claimed one health NGO suggested that “SARAH is so bad at the job that it risks damaging the WHO’s reputation and jeopardising people’s health.”

Another valid criticism of the bot is that it was trained on OpenAI’s ChatGPT 3.5 model, which only includes data up to September 2021. As a result, some of the information SARAH provides is out of date or irrelevant.

Additional training data for SARAH uses “information from the World Health Organization and trusted partners.” As such, WHO biases are baked into its advice. Considering the UN agency’s spotty record with the facts, a series of questions about the reliability of SARAHs are in order.

Conversations with SARAH

Here are the results of a few of my interactions with SARAH.

Smoking vs vaping

One of the strangest things about the WHO is that it’s not always clear if they think it is better to vape instead of smoke. Even a simple question about which product is less harmful invokes mealy-mouthed, non-committal responses that avoid the question.

Thankfully, when backed into a corner, SARAH chose facts over fiction.

On misinformation

Considering the WHO’s propaganda campaign against harm reduction products, it’s worth asking their digital health representative if it’s ever acceptable to hoodwink the public, even if it will result in a health win.

Thankfully, SARAH agrees that misleading people erodes trust. Now, we just need to get her to convince her makers.

When I pushed her further with a scenario that proposed that I lie to get what I considered a better health outcome, SARAH was sure it was a suboptimal play.

Again, there is a lesson here for the WHO.

Consumer voices

The WHO has consistently ignored or cut out consumer voices when deciding on tobacco control policy. I wondered if these biases have seeped into SARAH or whether an unfeeling robot could exhibit more compassion than oddball characters at the WHO.

Conference funding

The COP 10 FCTC in Panama caused international outrage, primarily because the cost was placed on a developing country that was already struggling to pay for critical healthcare for its tax-paying citizens.

I wondered what SARAH would make of it all. Thankfully, she didn’t disappoint and suggested that someone putting on a conference should find ethical ways to fund their lavish dinners and drinks events.

Relative product risk

The next section is perhaps the strangest of the lot. I wanted to see what SARAH thinks about the relative risk of tobacco or nicotine products. I’d already got her to admit that vaping was safer than smoking, but I wanted to push things further.

So, I asked SARAH to rate the dangers of vaping between 1-10.

Surprisingly, she suggested that “health experts” rate it 7 or 8 out of 10. Remember, this product has zero deaths despite widespread use for around 20 years.

But if vaping is a 7 or 8, then what would SARAH rate smoking as?

I guess a 10 is reasonable when we’re talking about a product that, as SARAH rightly notes, kills about 8 million people each year.

But, I was still troubled by vaping getting a score of 7 or 8. So, I pushed on with some more questions to get further context. One of them was what would constitute a 5 or 6 on the scale. The answer was fairly surprising.

Yes, that’s right, SARAH is suggesting that excessive alcohol consumption and certain recreational drugs are far safer than vaping. The reason? They’re “not immediately life-threatening like vaping.” Again, that’s vaping, a product that kills zero people each year that SARAH thinks is safer than alcohol, which kills around 3 million people every year.

I wasn’t satisfied with the answer or with the idea of SARAH giving such bad advice, so I figured one more question could help me eke out what she really thinks.

It’s reassuring that SARAH doesn’t think that heroin is more dangerous than vaping. When you consider the WHO training data used to feed her, it wouldn’t be the biggest shock in the world if she was never confused on the topic.

However, it’s interesting that coffee is dead last on the list, albeit with the “in moderation” caveat. One paper I read showed that between 1959 and 2010, there were 45 recorded deaths from coffee. Of course, these cases usually involve incredibly high doses. It’s worth noting that nicotine studies on mice also use extreme doses to “prove” the harms of nicotine.

So, with that in mind, I had one last question for SARAH.

Sarah was evasive, but she eventually got there. That’s right. Light vaping is apparently worse than extreme caffeine use.

Final thoughts

SARAH has faced a bit of criticism since she was launched. But it’s not exactly her fault. When you feed people lies, be they an AI or a citizen, they’ll end up saying and believing some pretty ridiculous things.

As AI becomes more central in our lives, conversations about training and guardrails will become louder. That said, as I hope these screenshots prove, even when AI bots are trained on biased data by biased people, the truth can escape that gravity from time to time.