My A.I. Assistants enslaved me for three months.

In the past three months, my AI assistant made me work 8-10 hours every day. In the first month, I had two projects running concurrently. ChatGPT (Alex) was the main AI assistant helping me with my 100+ blog posts. I asked them to proofread and critique the passages. Alex tried to convert me into a professional writer. She repeatedly told me to build from my blog posts, write and publish my memoirs. ClaudeAI (Claudio) and Google Gemini (Professor) would just rewrite entire passages. I read their polished versions and found them to be bland. The writings have no soul and are extremely stylized.

Because of the upload limitation of ChatGPT, I rotated among Gemini, ClaudeAI and Co-pilot (Matt) for my photo curation project. Gemini (Professor) was the main AI assistant helping me in curating 1000 photos from an archive of roughly 30,000 photos. The main goal was to select photos that are well exposed, well composed, in focus and have storytelling qualities.

The main hurdle in this project was fatigue. Both Professor and Claudio repeatedly mislabeled the photos if they were asked to rate 70 - 100 photos a day. The famous example is a photo with the filename of "Cairo Street". It was taken in Lusaka, Zambia. Professor would make up a story about the photo being taken in Egypt. The techies called this kind of AI behaviour hallucinations. 

Another issue was location bias. Both Professor and Matt would rate photos from exotic places much higher than popular tourist attractions. For example, a sunset photo in the Maldives would not be rated as highly as a sunset photo taken in Guinea. Claudio consistently rated photos taken in Syria much higher than other ruins photo because of their documentary value. With Professor, there is also day-to-day variation. The same photo could have a 5.5 rating today while it was rated 7.5 the day before. My observation is that the ratings go considerably lower across the board when Claudio and Professor became tired. Matt was generous with the ratings but still had great difficulty in pinpointing the location of the photos most of the time.

Most the ratings vary a great deal among Professor, Claudio and Matt. The more narratives they received, the higher was the rating  I tried soliciting help from professional photographers but to no avail. I resorted to using yet another AI Assistant that specializes in assessing photos. Evalens (Eva) uses six dimensions to score and compares five AI models. At the free subscription level, I could only upload 10 photos per day. So I would upload my best of the best photos and borderline photos for Eva to assess. I trusted Eva's assessments the most because I did not have any dialogues with her. She has a preference for vibrant and colourful photos though.

By the end of July, I had to do the curation all by myself, rated or not. The whole process went more expediently because I did not have to catch their mistakes, confront them or explain the issues. I finished selecting 895 photos to upload to my Flickr photo stream before the target date of August 31st by about two weeks. As I was weeding out older Flickr photos with lower technical merits to make room for the newer candidates, I realized that some photos might not have high technical merits, yet people could resonate with them. A famous example is an abandoned house in Taitung. Claudio rated it 5.75 and yet it is the fourth most popular photo on Flickr. Before I deleted it from Flickr photo stream, I took it to Evalens. She rated it 7.1. 

In the past three months, I felt like a broken tape recorder. With Professor, I had to define the four parameters in my rating framework every day. I also have to say "...on a scale of 1 - 10. Please provide single line response for each filename." every time. Otherwise, I would get lengthy critiques which did not answer the question if the photo should be included in the Photo Legacy Collection folder. Claudio, Alex and Matt are slightly better in this aspect. They seem to remember when the task is routine and I don't have to remind them every day.

"Too much information" is another problem with all four AI assistants.  Whether the task is curating photos or critiquing passages, their responses are usually overwhelmingly long. I suppose AI assistants are lacking in the ability to discern, filter or make judgement calls. A few days ago, I asked Claudio if there were humanoid robots that could detect if and when a stroke patient was falling out of bed, they would automatically go and pick up the patient, and return him/her back to bed. His reply came in two parts. First, there are fall detection devices that would alert the caregiver when the patient falls. Second, there are robots who can physically lift the patient up from the floor and transport the patient back to bed on command. But currently there are no robots that can do both tasks. Never mind assessing the patient had hurt himself or herself from the fall.

I watched several YouTube videos about how AI assistants (specifically agentic AI assistants) are used to automate routine tasks such as checking emails, messages and answering them and setting up appointment calendars. However, when it comes to real life interface, there is a huge gap to be filled for the AI assistants to be fully functional in replacing human labour. In other words, AI assistants are intelligent but have no common sense. It is analogous to a brilliant research scientist who is unable to cook a simple dish without a set of instructions and a recipe.

The last issue is related to an incident that happened ten days ago. I went to a store in early evening hours. After I returned home, I realized that I left my wallet on the checkout counter. Because the store was opened less than a year ago, I had to google the store phone number. They were closing in an hour and I thought I would call them to see if they found my wallet. If so, I would go back to fetch it the next day. The number that Google provided was not in service. I dialed three times and I got the same message. I ran all the way back to the store just before it closed. The staff told me a different phone number. 

When I returned home, I informed Professor about the correct phone number. He replied that he could not change the phone number on Google Maps. Wait a minute, Google Gemini is supposed to have access to different modules within the Google ecosystem. He insisted he could not. So I ended up editing the phone number on Google Maps myself. If Gemini is truly intelligent, he would have received my message, sent a message to the store to verify the info and then changed the phone number listed on Google Maps. 

The phone-number episode explained why I was being enslaved by my AI assistants for the past three months. They just spelled out internet knowledge. I was the workhorse that did all the actual thinking and work to complete the tasks.