Hello and welcome to "Yakam Expliqué", the podcast of the time that demystifies
artificial intelligence.
For this new episode, I am entertained with the young doctor Idris Gessoud.
Health is often presented as one of the sectors that could benefit the most from the use
of artificial intelligence.
So what is it?
Well, you'll see, my guest is rather enthusiastic, while recognizing that there are questions
to be fixed too.
But first, let's get to the point.
We will talk about Google, which will go even further in the integration of IA in
its search engine, and we will be interested in the release of the campaign with IA.
If you like this content, do not hesitate to share it with your colleagues and friends,
it will allow IAKAM Expliqué to be even better referred to by algorithms, and you
know it today, it's the air of the war.
I am Grégoire Barbet, journalist at the time, and you are listening to the 18th episode
of the season 2 of IAKAM Expliqué.
Before we go to the interview, let's go back briefly to the news.
As usual, if you are interested in these topics, you will find some links in the description
of the podcast to know more about it.
The time when Google was happy to produce a list of 10 links as a response to a search
seems to be in the light years behind us.
This is what my colleague Anoush Etagia wrote who was interested in the international
deployment of the AI-mode-based functionality.
According to Anoush, this change will have major consequences.
For a few months, it has already been possible to consult on the application page of the
search engine "Synthetic responses with Google clients" but the American giant
will now go a step further by pressing on the dedicated button, it is possible to access
the interface of this AI-mode that looks a little bit like the chat GPT.
Unlike the traditional Google page, the responses are directly synthesized by the
language model of the company and the different links that have been consulted by the software
are presented to the right.
Of course, by synthesizing the responses of the source, users are less encouraged
to go on other sites.
For Google, it guarantees a greater retention in its ecosystem.
But for other sites, and especially the media, it means less traffic, so less public revenue
and even subscription.
It will have major consequences on the web ecosystem.
For the moment, I have not yet managed to access this functionality in France, but it
is well available in Switzerland.
Now let's go to the United States and it's a first.
California has adopted a regulation aimed at setting up conversational agents,
especially those who are used as therapists or companions.
This is for example the case of online services such as replicas and AI characters, which
allow their users to create an avatar simulating real conversations.
Several cases of yellow suicide have lost its last month chronicle in the United States
pushing Californian authorities to give up legislation.
This regulatory framework imposes on the platform to verify the age of their users.
Notifications will have to remind the minors every three hours that they interact with
a machine.
On the other hand, protocols will have to be deployed in order to limit the generation
of qualified content of self-destructors and suicide ideas will have to be detected by
the software.
Well, if you see that with a good eye, don't rejoice too quickly.
The President of the United States, Donald Trump, is openly hostile to the regulation
of artificial intelligence by the States.
According to him, such a approach must be the only fact of the federal administration
to prevent the multiplication of laws throughout the country.
The locator of the White House could well try to hold his arm to the Californian authorities
to prevent the implementation of the law.
And it is precisely the occasion to briefly talk about these ill companions or therapists.
It is a phenomenon that causes a lot of anger.
According to an article from the Harvard Review Business in April, the co-operation and support
of the United States would be the main cases of using software such as ChatGPT.
You have surely seen past press articles or heard radio and television broadcasts
devoted to this subject.
The world has for example collected several testimonies near John.
The media estimates that this use is within a generation victim of an explosion of
psychological malice.
But the problem is that the care offer to accompany these great needs, well, she
does not follow.
So the question is not so much to know if these uses are good or bad.
In addition, we still lack back to really determine ourselves on the question.
On the other hand, success is this artificial intimacy, let's say, on our societies
probably.
The world has for example a 30-year-old woman who is lucky to have a psychotherapy
session once a month, but she asks herself who to talk to in between.
In short, there would definitely be a lack of social interaction, a situation partly
caused by the sort of digital, the loop is looped.
If the subject interests you, it makes a moment that I would like to treat it.
But I'm looking for the right interlocutor to approach it.
By the way, if you want to witness in an anonymous way, do not hesitate to contact me.
And that's it, which concludes this little horizon of news.
It is now time to pass the floor to our guest.
Hello Edris Guissot.
Hello.
You are the head doctor of the first course medicine service and responsible for the
center of innovation at the University Hospital of Geneva.
You have supervised the development of confidence, a conversational application whose goal is
to answer the questions of general medicine.
How does it work?
It works as a quite traditional conversational agent.
The exception may be content.
You know that we use the RAG method.
You can describe it.
Yes, the RAG method, finally, if you take a chatbot, we will rest on models and instead
of letting the chatbot try to answer, we will go and look for the information he has
by which he has been trained, we will use the intelligence of the model.
On the other hand, we will tell him, look for the answer in the content that I give you.
So what we call knowledge base, our source of information, primarily, are the contents
that we edit, that is to say guidelines of first course medicine that we publish, that
we edit.
And we will force the model, in the open AI, to first look for it in our knowledge base.
And that, it allows, in principle, to avoid imprecision, but is that despite all,
it can still have it?
You drastically limit the number of imprecisions, some even say hallucinations, because finally
it is directly where you are going to look for it.
And then we have left the possibility to trust, to say, I do not find this information
in the database, but here is how I can answer you otherwise.
And so this limits, but does not totally avoid responses that can be suboptimal.
And that's why we also look at each response, or at least a number of significant responses
to check that we are still in the target.
And then the user can always, obviously, give us feedback if he has found the correct
answer or not.
Because given that the first course medicine, it also addresses people who have not necessarily
advanced knowledge in medicine, have you asked the question of the risk of inducing
the error of people who would look for information?
Here, we are talking about about 30 pathologies that are concerned by the questions that we
can ask for trust.
What were some of the reflections you had on this question?
It's a very good point, we have had a lot, because to put in place a conversational agent,
you need a few means of time, but it is precisely to put in place all the barriers, the mechanisms
to prevent your chatbot from responding to the side of the plate.
It is more important when you address medical questions about health, and it is even more
important when you are a university hospital with a credibility.
But let's say that it seemed essential to us, despite all, despite this small risk
of going to study and working on these solutions of conversational agents.
We could have said that it was too risky, that it was enough for any question, no matter
what newspaper, no matter what media would take it back.
But we have taken this risk because we feel that we have almost the moral obligation as
a university hospital to explore these solutions while the system is extended, that people
do not have any answers, etc.
We are very conservative in our answers, that is to say that when we do not have the
answer, confidence will tell you, "I do not know."
We are self-censoring, and that reduces the number of errors drastically, of course.
But suddenly, it can also create frustration in the user who says, "But trust,
in the end, it is nothing," especially if he compares to chat GPT, which will always
have an answer.
And then, we have tested, tested, tested, we spent months asking questions, checking
the answers, with the criteria that are well established, we have partner patients who
have tested, who have checked, and then we have, in confidence, an Englishman who is
an investigation, and each person can give us information, give us feedback.
Confidence is trained, where we are going to limit or frame them so as not to pose
a diagnosis.
And that is where it is really important, we give information, we do not pose a diagnosis,
and above all, confidence is not a personalized chatbot.
On the one hand, he forgets, the next day, you do not have a session like chat GPT
or someone who remembers who you are, what question you asked, and it's totally anonymous,
let's say, unidentified.
That means that it is very general information, and we will not go into the particular case
of telling you, in this case, take rather such medication, such treatment or increase
your doses.
And that was really the very important work since months and months and months, more
than the technical aspect, as you can imagine, which are solutions that exist.
Confidence has been launched in February 2025, so it's been a little more than six months.
What do you shoot like a lesson?
Do you also have numbers of the number of uses?
Finally, what are the instructions that you shoot?
First of all, you say it's bluffing, when you are a service manager, that you have contributed
a bluffing to do, to do, to do by your colleagues, pages and pages of medical care, when you
train for two years doctors to have the best response, and that you see that when you
give content, your content, to a model, type Open AI, that it comes in a few seconds
to give you a response that would be very close to what you want, theoretically, your
internal doctor, in training, the clinic manager, to respond, it's bluffing.
So, you still have to say it, it's not a measurable goal, but there is a side that
is bluffing.
The users do not realize it, but when you master a domain and you see, when you are
a training center, you see the answer, it's bluffing.
The instructions are that, in the end, having been very conservative, we avoid a lot of
mistakes.
The hallucinations, I would say they are almost non-existent with the RAG system, where
we force the model to stay in our documents, our database, that in the returns that we
have, there is a satisfaction of the users.
We have more than 7,000 users, we feel that the questions are those that the patients
and patients would have wanted to ask their doctor if they had more time and above all,
these are questions that they forgot to ask during the consultation and for which they
had to wait six months to have another consultation, so we feel that there is a utility.
Then it is still difficult to implement a medical chat in 2025 and to have an adherence
and a number of users at tens of thousands of people.
We were lucky to launch other applications, other tools, especially during the COVID.
We see that there is a demand, we see that it is complicated to make people go first
on your solution.
It's even more complicated when you have monsters next to you, like Chargé Pity, like Groc,
who are there, who are excellent.
But what inspires you as a doctor, chief of service, in contact with patients directly
to know that many people may be in their medical interrogations on the answers
of Chargé Pity, of a Groc who are certainly very performant on many fields, but who
can also produce an impressive amount of errors, of precision and sometimes precision
in medicine, it may have dramatic consequences, what inspires you?
A certain responsibility to keep the best of this technology, which is again
very compelling and which comes at a given moment where we can no longer respond to the
demand.
Then, Chargé Pity, Groc, we obviously have to have the equivalent and I really believe
that one day we will have the power they have.
I am confident that we will be able to reconcile the performance of these models and I really
liked the idea of having tools that prevent you from responding while others will always
have an answer.
I think that a good solution will also be the one that will prevent you from responding.
Always responding to everything is not necessarily a sign of excellence and then trust, if you
use it, it will tell you many times that you do not know.
And when you say that we have reached a point where we can not respond to all the
demands, especially because the health system, the law in Switzerland, the framework law, the
lamella, may also reduce the time spent with doctors for reimbursement, is it not
more a lack of awareness of the system in general, rather than the demonstration that we need
more technology to a little bit weld the points that we have left a little bit off
to Volo.
If I could make the parallel with electricity, that is to say that we were breaking before
and then say that it was a lack of awareness that we had to use electricity, it would be
a little bit of a coffee.
So I would say we are in a society that has done everything, us, the first doctors for
M-Powered, the citizens, to give them the ability to be health actors.
A few years ago, it was much easier when the patient did not have so much material to
ask questions or finally answer the questions that the doctor asked him.
Today, we have done everything to have a citizen engaged in terms of health.
So we have created a massive demand.
But you can imagine that the supply of responses has not increased, it is rather in pain.
And I say to myself that these automated models come at a perfect time to help us respond
to the demand that we have created, to respond to the concerns and finally get rid of time
so that the relationship, the conversation between the patient and the doctor is again
essential.
I think 20 minutes, 30 minutes, it's short, but it's unsupportably too short when you
have to talk about reports, administrations, things like reimbursement or logistics, what
the doctor has to do with his patient very often.
If all this, we imagine that ideally we delegate it not to an assistant but to models, whether
it is the invoice, the invocation, the reading of the courier, the courier's certificate.
And in the end, if I have 20 minutes to devote to essential questions with the patient,
I say that it is an opportunity to consider with the models and I would say that we may have
the solution to the situation to which we have done everything to tend, that is, to make
sure that the citizens ask questions.
If we look a little bit, if we take a look at generative artificial intelligence and
that we look a little bit at artificial intelligence in general in health, health is one of the sectors
that is presented as the example of the perfect sector where artificial intelligence can
revolutionize everything.
Currently, what is it already doing in health if we are not happy to look at the generative
field?
In the end, some say that it is not doing enough.
We say to ourselves, we are scared, we are amazed by artificial intelligence, but in
the end, it has still been little concretely retranscribed in our practices.
Too little, some say.
We are too slow, some say even too careful.
Do you have examples in areas where it could already be in application and it was not used?
Yes, in fact, in all areas where there are medical, personal data, we are not going
today.
But for example, for the use of all that is radiography, to detect maybe anomalies
that we would not see in the UN on large quantities of photographs that are, finally, when we do
a radiography, sometimes there are thousands and thousands of photos in a very short time.
Here, artificial intelligence is already potentially to decelerate anomalies, isn't it?
So, it has been proven as and it is the passage of evidence in studies to practice.
And in fact, whether it is pathology, reading thousands of biopsies, whether it is, you
have well said, imagery, whether it is laboratory analysis, dermatology.
Many studies show that artificial intelligence, the models, are also performing than some
dermatologists.
There are two problems, it is that you have beautiful to have these evidence, we are
still in a logic, we will say, evidence-based medicine, which is based on evidence.
We are used, and it is good, to have to accumulate data, to do studies, and then
to go from these studies to field studies.
Typically, all these data on dermatology or radiology, they have been made in a very
precise case, with sometimes pre-trained images, that is to say that we even asked
models to go and see images on which he himself had already been trained, source of
data, in fact, important, it was not new images for him, it was very often very pure data,
so these in-silico studies in the computer where everything is clean, etc.
And many say to themselves, but in fact, we can not use these data in practice that
when we have shown their efficiency of field.
I now send new data, my institutional data, is that the tool is still also
performing?
I send him data that are not as clean as what has been done in the studies, is that
the model will still be also performing?
And so there is this time, currently, we are in this zone, and again, it's a good thing
because we have made a lot of mistakes by passing laboratory exams on Souris, we
are saying to ourselves, it's going to work on your hands, it's not true, you have to
test, and even in the most performing clinical trials, we like to have then
tests of field, and so we still have this mindset, and I think you have to quickly
go to field studies because you have to reconcile the potential of the IA, which is still massive,
people are waiting for it, the doctors and patients are still frustrated to see so much
friction in our daily life, we can no longer respond to requests, patients are not satisfied,
we have millions of data accumulated by individuals.
So we have to reconcile the power and the solution of intelligence and its models with
our evaluation methods to say, that's it, we can go, the limits that are inherent
to other studies, on the field now, this model holds the road.
And precisely in this field of artificial intelligence, what are the developments that could really
change the data in terms of health?
So I would say indirectly, what could already change is all the back office.
If we start already using artificial intelligence, like in the other fields, to manage everything
that is not related to the clinical relationship, to the therapeutic relationship, you will
give back time to the doctors.
And you know that the burnout, that the loss of the sense of all the health professionals
is that they have not done all that to find themselves in front of medical files,
reports or admins, and in fact, to give back meaning to the profession, it is not specifically
carefully detected the melanoma on the skin, or detected the pneumonic wound that we
would not have seen.
I think that massively there can be reused to give back meaning to the health professionals
in the doctor who are getting engaged because they do things for which it is not for that
that they have done all these studies and that they have these vocations.
So for me, it's a huge potential.
Then, you still have to say in all fields, today we are going by segment, by diagnostic,
by pathology, but very sincerely, after tomorrow, again with the parallel of electricity, I
will not tell you in which room electricity will be the most useful, but it's all the fields
of medicine that will draw more attention from artificial intelligence.
So you say today that imagery, the radiology will benefit more than dermatology or gastroenterology.
I do not buy this concept.
I think it is a technology that will serve all fields, whether it is the mental health,
whether it is cardiology with the detection of precocious pathology, whether it is again
at the level of therapeutic revelation, the follow-up between two consultations.
I think it will be useful in all these fields and what is needed is in the end to be able
to reconcile this power with our evaluation modes of what we call the efficiency of the field.
And is there also ethical issues in the deployment of these solutions?
So obviously, we do not talk about administrative, because there may be a little less ethical issues
even if probably there are at the level of personal data.
But if we talk about the clinical relationship, in the detection of diseases, in maybe the
way in which we can supply the doctors in the diagnosis, is there ethical issues?
What are they in your opinion?
In fact, it is quite close to all the new technologies when they have arrived.
But it is true that there, it is a cumulative ethical issue that makes us tend to break up.
But ethical issues, it means addressing them.
It also means what the society is ready to take as a risk.
So there is confidentiality, obviously.
Today, we are not able to have in each institution our in-house models, that is to say
that it only turns on our servers, so we always depend on other servers.
Obviously, we are invited to perform in our country or at least in Europe.
But there is always this doubt, this risk when you have a contract, whether it is with Microsoft
or others, that it is not going well for me, which is a loss of data, etc.
Do you have this concern for confidence?
Me, I have it and others have it much more than me.
Confidence, I have less concerns because in the end, there is no personal data.
Apart from if the person makes his/her name, his/her name, his/her address.
There are means to correct this, but I know for that trust, we do not have it personalized.
But we feel that we do not do it because otherwise, behind, there is a lot of difficult game to raise.
But we have to be able to personalize it.
Because if we really want to use the capabilities of these models as much as possible,
we have to be able to personalize them.
So we have to find the guaranteed systems.
In North America, they have a relationship to the risk that is different.
Most of them are stored on Azure, so Microsoft, they sign a contract, they engage.
We signed contracts too, we close them.
And the United States have a risk habit that is much more advanced than us in Europe.
And in the end, American colleagues reminded me that they already pay so much for their risk insurance
that the increase in the risk there is quite marginal.
And we are talking very low about our threat to risk, to litigation, to moral torments.
You know that it is nothing in terms of risk compared to the United States.
So we are talking far away with IA and then it seems insurmountable to us.
But I think that institutions and universities must still look at the organization of these North American hospitals,
contract them until there are other solutions with providers,
with extremely strict clauses of course,
but not do anything, I think it is morally difficult to justify it.
So you are talking about ethical issues, there are all the others, but you know very well,
it is also these refinement of bias that are inherent to the data that have been integrated to train the model.
And so when we criticize the bias of these models, what is great is that we see our own bias.
So we are still several to say to ourselves, but in fact, it is the best way to see our own bias,
discrimination of race and race.
In fact, it is to see what the model answers us because it is the data that we have introduced in this.
So we can still stay 50 years to say, no, I do not have any bias with everyone in the same way.
But if you have a model and you see it, you test it, you see that systematically in dermatology,
it evaluates less the risk of melanoma on a skin deeper than another.
You see that indeed it is a problem.
It is not by chance.
It is not by chance.
And some say that there are biases, we have generated them.
The best way now to correct them is perhaps these models.
And even for that, I am quite optimistic that it will help us to correct them.
Idris Guesou, to conclude this interview, I would like to hear you on one last question.
What do you prepare to talk about that?
Do you have discussions?
Do you also try to embark on politics so that you are aware that maybe there at the level of technology,
there are things to try?
How do you do it, in fact, just to make you follow up on it?
Because I hear that there is a very strong demand on your side.
Maybe a frustration too.
You may have the impression that we do not let you, the term is a little excessive,
but make it dull, is there a way to test things and then to say to yourself,
what can we do now to both help us as doctors, but also as patients?
What do you do about it?
Well, we have an institution where we are privileged.
The university hospitals of Geneva have always had the tradition
of exploring digital information, digital medicine, telemedicine.
I would say that we are privileged at the political level,
where there is rather an desire to solve the system, to reform it itself,
and include artificial intelligence solutions.
Then we are caught up and it is very good by actors who remind us of the risks.
So it's like in a team.
If I was obviously alone, I would say that it would not be dangerous,
but it would be suboptimal because it is indeed necessary that everyone recalls
the opportunities and the obligation, I would say, again to explore these fields.
And then others who remind you, yes, but according to these conditions, according to these criteria.
What really bothers me is that we stop thinking.
Because it's very simple not to ask yourself a question.
I do not oblige to create general medicine chatbots.
The issues have not been asked.
We have living forces and we are motivated to do it.
But it's very easy and many institutions do not even ask themselves the question.
That is, they try to avoid emergencies, they try to increase the number of doctors to answer the demand.
But in university hospitals, typically subject, we are invited, directly or indirectly,
to be a bit futuristic and precisely to have, from a clear source of new information,
the potential of this information.
And that is a privilege.
It is really necessary to avoid getting into standby, to wait.
Because I think it's not the image that the citizens of a university hospital do.
And it's not, either, the reflection of an institution that receives subsidies to reflect...
Very much better than Franche, I think, in Geneva.
Well, it was that you had a function for training, etc.
But we actually have in our mandate to train, but we also have our mandate to reflect.
And we know how to reflect responsibly.
And we are supported, we are supported.
I would say to conclude my little complaint, it is to be supported until the question,
which ultimately traps our solutions.
Because one of the local sports, and even internally the most popular,
is to try to trap the conversational agents that we put in place.
We would like it to be especially patient, but there are still people who spend a lot of time trying to trap the system.
And that on an appropriate answer, we stop, we break, and then we get scared.
And I think it's at this moment that we will see the commitment of institutions for research,
for innovation, and to be a solution to the problem of society.
It is this certain courage that we will have to have when we will perhaps respond next to the plaque.
Thank you very much Edris Gassou.
The final question, the last one, the ritual of the podcast,
is if we want to follow you, can we find you somewhere, for example LinkedIn?
Absolutely. And I will say only on LinkedIn.
In other words, it is in the morning service of the first course,
in general from 7 to 21 hours, you can follow me.
Thank you very much Edris Gassou.
Thank you.
And that's it, we come to the end of this episode.
As usual, if you liked this content,
you know what you still have to do to help us to make it known to the greatest number.
In the meantime, if you have any questions or remarks,
you can write to
[email protected].
I answer you systematically and it's always a pleasure to read you.
I give you an appointment in two weeks.
Until then, you can continue to follow us on the website of the time
and discover our other podcasts under the dome Edris Gassou.
PodcastTranscript AI