I am also curious about the good results you're getting from AI in the classroom. I've had to ban using AI in searches for information during class discussion (I used to love having students look up answers for themselves on search engines) because the answers are wrong--and wrong in unique, bespoke ways for each student, even when identical prompts have been used in ChatGPT. (These are answers to non-controversial fact questions, like "What is the concentration of methicillin that will kill these susceptible bacteria?")
ChatGPT also fabricates references, even when students ask ChatGPT to simply reformat an existing correct reference page into another style. The fabricated references look real, but are amalgamations of author names and article titles form a variety of places, with publication dates, volume, page numbers and DOIs entirely fabricated. Colleagues of mine in university libraries also note that they are seeing the same thing across academic fields.
So, uh, for me personally as a professional woman, I am less likely to use AI because it doesn't work. It makes more work for me because it's so often wrong.
Use Perplexity.ai instead. It lists the references used in its answer, which the user can then check. Also have users experiment with using different kinds of phrasing in their question (prompt) to see how the answer changes.
Thank you for the tip about Perplexity. I will have to try it.
On ChatGPT, the prompt was identical for each of the students who were using it; the answers returned by ChatGPT were not. This is speculative, but perhaps the answers were influence by previous ChatGPT prompts or previous search history? A friend of mine who has very different opinions than I do on a particular medical topic and I put identical prompts into ChatGPT, and we each got our own bespoke answers that comported with our opinions. Certainly, in ChatGPT responses to later prompts in a thread are tied to the earlier prompts.
For instance, a student who had her References flagged because they were, ahem, all fake showed me her process and ChatGPT logs. She had a list of real references that were correctly formatted in MLA style. She had asked ChatGPT to reformat this document into APA style. This it did not do correctly. In this project, the student had been in correspondence with an instructor who replied that her references were not in APA style. When the student would go back to ChatGPT with the query "my prof says [citation] is not APA style" (I know, not the best prompt), ChatGPT would reply "sorry. Here is the corrected citation" and then add more random items to the newly generated citation. In one case, a reference for a paper that in fact had seven authors just kept having fabricated author names added to it by ChatGPT until it had 25 authors. The volume number, page numbers, publication date, and DOI changed with each iteration through ChatGPT. This is interesting to me, since more than 20 years ago I was using EndNote for this same task, which it did flawlessly.
Perhaps Openai is adding its log of a user's previous questions to the prompt for the current question? Perhaps as a move to improve UX? This could change the output considerably for identical questions posed.
I stopped using Chat GPT about a year ago. I'm not from the USA, whereas ChatGPT has very strong guardrails that constrain its response to those that are acceptable in the USA. This affects both the language and the ideology of its responses.
This also applies to Perplexity, but less so. My biggest frustration with Perplexity is the limited variety of references it derives its answers from. Don't expect to get good reporting of cutting edge science, for example, though it will cast a somewhat wider net if carefully prompted.
I think you are being unduly negative. Perplexity is good for confirmatory and for more precise info than you have in one's memory ie. a reliable enough reporter for non contentious issues that you don't need to"research".
In journalism there are no "non-contentious issues not in need of research". Nothing is like it seems. Journalists should be fleas in the fur, always doubting what they are told, always looking for the other side..
I do not need an AI to tell me what I already know/can quickly research. When it starts to challenge me, points me in other directions, that is when it becomes useful. Till then I will stick with my always contrary husband. He also makes better coffee.
No, reporting. Maybe good journalism is in part advocacy for the masses.
I find in newspapers nowadays too much regurgitating of what is officially told. The "other voices" are now found in Substacks. Only heard by an elite group.
EDIT: I have to add that AI used in specialistic settings, e.g. tge medical field, can better the performance of the users. But those systems are built and "taught" by experts in their field. Not loaded by scraping the Internet.
Agree with the AI sceptics here. My trials so far haven't convinced me that it's worth it. I don't trust the results and what I get is no better than a Wikipedia page on the topic. And then I found out how much more energy and water is being used. I used to embrace all new tech but now I am a bit older and can take a more sceptical view as I can see it's not all positive.
In my opinion, Wikipedia is much better. At least if you look at a Wikipedia page and I look at the same Wikipedia page we are getting the same information, not whatever AI "thinks" we "want" to read.
I think it's unsurprising that high-achieving women, especially in an educational setting, are less likely to use AI for tasks. I think a lot of us are very sensitive to fears of being accused of not being good enough/doing our own work. Even in more egalitarian countries, narratives about how high achieving women "really" got their positions are rampant. Anything that would add to that narrative--like, that I as a woman did my work using AI--is to be avoided.
I also think that this is a result of women being used to doing more tasks for themselves, both at home and in the workplace. In every academic workplace I have ever worked in, female teachers--but not male teachers--make their own copies and coffee, and wash their own dishes in the break room. Long ago, when typing was a class students could take in high school or business school, my mother taught it: First, to female students who were studying to be secretaries to businessmen, and later to female students who were going into business themselves (and would be expected to type their own correspondence, make their own appointments, etc.) Paradoxically, I think people who are used to doing more tasks would be less likely to use AI because using AI is less like doing a task yourself--it's more like offloading it onto an assistant. I think that there's deep cultural training that makes women less likely than men to jump at offloading a task to an assistant.
Thank you Alice Evans for these interesting insights. One of the studies you mention found out that the highest achievers among female students are likely to be the most reluctant ones to using genAI. That's something to further examine. I've seen the same reluctance among some high achieving female colleagues, too. They are disappointed with the results of current genAI, which is a fair point :). I am typically trying to explain that, despite the current 'poor' results of current AI tools, genAI is like using the email when it first appeared. It's a productivity tool that will integrate many professions, whether one likes it or not. Today, it requires practice to get to better results.
I am also curious about the good results you're getting from AI in the classroom. I've had to ban using AI in searches for information during class discussion (I used to love having students look up answers for themselves on search engines) because the answers are wrong--and wrong in unique, bespoke ways for each student, even when identical prompts have been used in ChatGPT. (These are answers to non-controversial fact questions, like "What is the concentration of methicillin that will kill these susceptible bacteria?")
ChatGPT also fabricates references, even when students ask ChatGPT to simply reformat an existing correct reference page into another style. The fabricated references look real, but are amalgamations of author names and article titles form a variety of places, with publication dates, volume, page numbers and DOIs entirely fabricated. Colleagues of mine in university libraries also note that they are seeing the same thing across academic fields.
So, uh, for me personally as a professional woman, I am less likely to use AI because it doesn't work. It makes more work for me because it's so often wrong.
Use Perplexity.ai instead. It lists the references used in its answer, which the user can then check. Also have users experiment with using different kinds of phrasing in their question (prompt) to see how the answer changes.
Thank you for the tip about Perplexity. I will have to try it.
On ChatGPT, the prompt was identical for each of the students who were using it; the answers returned by ChatGPT were not. This is speculative, but perhaps the answers were influence by previous ChatGPT prompts or previous search history? A friend of mine who has very different opinions than I do on a particular medical topic and I put identical prompts into ChatGPT, and we each got our own bespoke answers that comported with our opinions. Certainly, in ChatGPT responses to later prompts in a thread are tied to the earlier prompts.
For instance, a student who had her References flagged because they were, ahem, all fake showed me her process and ChatGPT logs. She had a list of real references that were correctly formatted in MLA style. She had asked ChatGPT to reformat this document into APA style. This it did not do correctly. In this project, the student had been in correspondence with an instructor who replied that her references were not in APA style. When the student would go back to ChatGPT with the query "my prof says [citation] is not APA style" (I know, not the best prompt), ChatGPT would reply "sorry. Here is the corrected citation" and then add more random items to the newly generated citation. In one case, a reference for a paper that in fact had seven authors just kept having fabricated author names added to it by ChatGPT until it had 25 authors. The volume number, page numbers, publication date, and DOI changed with each iteration through ChatGPT. This is interesting to me, since more than 20 years ago I was using EndNote for this same task, which it did flawlessly.
Perhaps Openai is adding its log of a user's previous questions to the prompt for the current question? Perhaps as a move to improve UX? This could change the output considerably for identical questions posed.
I stopped using Chat GPT about a year ago. I'm not from the USA, whereas ChatGPT has very strong guardrails that constrain its response to those that are acceptable in the USA. This affects both the language and the ideology of its responses.
This also applies to Perplexity, but less so. My biggest frustration with Perplexity is the limited variety of references it derives its answers from. Don't expect to get good reporting of cutting edge science, for example, though it will cast a somewhat wider net if carefully prompted.
In the time it takes me to check the Perplexity answer I can do the research myself.
I think you are being unduly negative. Perplexity is good for confirmatory and for more precise info than you have in one's memory ie. a reliable enough reporter for non contentious issues that you don't need to"research".
In journalism there are no "non-contentious issues not in need of research". Nothing is like it seems. Journalists should be fleas in the fur, always doubting what they are told, always looking for the other side..
I do not need an AI to tell me what I already know/can quickly research. When it starts to challenge me, points me in other directions, that is when it becomes useful. Till then I will stick with my always contrary husband. He also makes better coffee.
"Nothing is like it seems." . . . . You seem to be referring to advocacy rather than reporting.
No, reporting. Maybe good journalism is in part advocacy for the masses.
I find in newspapers nowadays too much regurgitating of what is officially told. The "other voices" are now found in Substacks. Only heard by an elite group.
EDIT: I have to add that AI used in specialistic settings, e.g. tge medical field, can better the performance of the users. But those systems are built and "taught" by experts in their field. Not loaded by scraping the Internet.
Maybe women have more common sense, and realize AI is hyped beyond its current possibilities.
Agree with the AI sceptics here. My trials so far haven't convinced me that it's worth it. I don't trust the results and what I get is no better than a Wikipedia page on the topic. And then I found out how much more energy and water is being used. I used to embrace all new tech but now I am a bit older and can take a more sceptical view as I can see it's not all positive.
In my opinion, Wikipedia is much better. At least if you look at a Wikipedia page and I look at the same Wikipedia page we are getting the same information, not whatever AI "thinks" we "want" to read.
I think it's unsurprising that high-achieving women, especially in an educational setting, are less likely to use AI for tasks. I think a lot of us are very sensitive to fears of being accused of not being good enough/doing our own work. Even in more egalitarian countries, narratives about how high achieving women "really" got their positions are rampant. Anything that would add to that narrative--like, that I as a woman did my work using AI--is to be avoided.
I also think that this is a result of women being used to doing more tasks for themselves, both at home and in the workplace. In every academic workplace I have ever worked in, female teachers--but not male teachers--make their own copies and coffee, and wash their own dishes in the break room. Long ago, when typing was a class students could take in high school or business school, my mother taught it: First, to female students who were studying to be secretaries to businessmen, and later to female students who were going into business themselves (and would be expected to type their own correspondence, make their own appointments, etc.) Paradoxically, I think people who are used to doing more tasks would be less likely to use AI because using AI is less like doing a task yourself--it's more like offloading it onto an assistant. I think that there's deep cultural training that makes women less likely than men to jump at offloading a task to an assistant.
"If you want a job done well, do it yourself."
Thank you Alice Evans for these interesting insights. One of the studies you mention found out that the highest achievers among female students are likely to be the most reluctant ones to using genAI. That's something to further examine. I've seen the same reluctance among some high achieving female colleagues, too. They are disappointed with the results of current genAI, which is a fair point :). I am typically trying to explain that, despite the current 'poor' results of current AI tools, genAI is like using the email when it first appeared. It's a productivity tool that will integrate many professions, whether one likes it or not. Today, it requires practice to get to better results.