Generative artificial intelligence took over my life.
For one week, it told me what to eat, what to wear and what to do with my kids. It chose my haircut and what colour to paint my office. It told my husband that it was OK to go golfing, in a lovey-dovey text that he immediately knew I had not written.
Generative AI, which can spin up research reports, draft emails and converse like a human being based on patterns learned from enormous data sets, is being widely adopted by industries from medicine to business consulting as a timesaving tool. It’s popping up in widely used consumer apps, including Siri and Alexa. I conducted this admittedly ridiculous experiment to see how its spread might affect the largest workforce of them all: harried parents.
In all, I used two dozen generative AI tools for daily tasks and nearly 100 decisions over the course of the week. Chief among my helpers were the chatbots that every big tech company released in the wake of ChatGPT. My automated advisers saved me time and alleviated the burden of constantly making choices, but they seemed to have an agenda: turn me into a Basic B.
Spark in charge
I told the chatbots that I was a journalist conducting an experiment and that I had a family, but not much more. AI’s first task was to plan our meals and generate a shopping list for the week, and within seconds, it was done.
Hill: You will decide what my family eats and what I cook this week. I’m going grocery shopping now. What should I get? Make the list ordered by section, starting with produce
ChatGPT: Here’s your grocery list, organised by section:
Produce:
Fruits:
- Apples (3-4)
- Bananas (one bunch)
- Oranges (four-five)
- Berries (Strawberries or Blueberries, one pint)
- Lemons (two)
Vegetables:
- Baby spinach (1 bag)
- Carrots (1 bag)
- Bell peppers (2 red, 2 green)
- Broccoli (2 heads)
- Sweet potatoes (4-5)
- Onions (yellow, 2)
- Garlic (1 bulb)
- Avocados (2-3)
- Cherry tomatoes (1 pint)
- Mushrooms (1 pack)
Dairy:
- Milk (1 gallon)
- Eggs (1 dozen)
- Cheddar cheese (1 block)
- Greek yoghurt (1 large tub)
- Butter (1 pack)
Meat/Poultry/Fish:
- Chicken Breasts (2 lbs)
- Ground turkey (1lb)
- Salmon fillets (2)
- Bacon (1 pack)
Grains/Pasta:
- Whole Wheat Bread (1 loaf)
- Quinoa (1 box)
- Brown rice (1 bag)
- Whole Wheat Pasta (1 box)
Canned Goods:
- Black Beans (2 cans)
- Diced tomatoes (two cans)
- Tomato paste (one can)
- Coconut milk (1 can)
Condiments/Spices:
- Olive oil (1 bottle)
- Balsamic vinegar (1 bottle)
- Soy sauce (1 bottle)
- Honey (1 jar)
- Salt and pepper
- Italian seasoning
- Paprika
- Cumin
Frozen foods:
- Mixed vegetables (1 bag)
- Frozen peas (1 bag)
- Frozen berries (for smoothies)
Snacks:
- Almonds (1 bag)
- Dark chocolate (1 bar)
- Whole wheat crackers (1 box)
- Hummus (1 tub)
Beverages:
- Coffee (ground or beans)
- Green Tea (1 box)
Hill: Did we forget anything? (With image of grocery cart attached.)
ChatGPT: Based on the image, it looks like you have most of the key items from your list, including produce, dairy, and some pantry essentials. However, I can’t see everything in the cart, so here are a few things to double-check:
Meat/Poultry/Fish: Make sure you have chicken breasts, ground turkey, and salmon fillets.
Grains/Pasta: Ensure you’ve got whole wheat bread, quinoa, brown rice and pasta.
Condiments/Spices: Check for olive oil, balsamic vinegar, soy sauce and the necessary spices like salt, pepper, Italian seasoning, paprika and cumin.
Snacks: Almonds, dark chocolate, whole wheat crackers and hummus.
If you’ve got these items, you should be all set for the week! If any are missing, you might want to grab them before you finish checking out.
My daughters were enchanted by the disembodied voice with infinite patience for their questions. (There are nine voices to choose from; I went with an upbeat male one.)
They decided it should have a name. It listened in as they chattered away with suggestions – many scatological, because they are 4 and 7 – and then chimed in:
“How about the name Spark? It’s fun and bright, just like your energy!”
The only junk food on Spark’s grocery list was a bar of dark chocolate. AI’s daily plans involved an hour of exercise, stretch breaks during work hours and cooking elaborate meals three times a day.
It planned family games in the evenings, including Pass the Story, in which we and Spark took turns telling a tale the chatbot started about “a towering tree deep in an enchanted forest”. The AI-optimised week felt like a wellness retreat.
My AI handlers didn’t just want me to survive the week; they wanted me to thrive.
On a Zoom call, OpenAI’s Joanne Jang, who is responsible for how ChatGPT interacts with users, said model behaviour was still an “ongoing science”. One of OpenAI’s primary instructions to ChatGPT, though, is to “benefit humanity”.
Relief from decision fatigue
I’d been wanting to repaint my home office for more than a year, but couldn’t choose a colour, so I provided a photo of the room to the chatbots, as well as to an AI remodelling app. “Taupe” was their top suggestion, followed by sage and terra cotta.
In the Lowe’s paint section, confronted with every conceivable hue of sage, I took a photo, asked ChatGPT to pick for me and then bought five different samples.
Hill: Which of these shades is closest to the office paint colour we decided on?
ChatGPT: Based on the olive green shade we discussed earlier, here are a few swatches from your image that seem closest:
- Snipped Chive (8002-30C): This is a rich olive green that likely matches well with the shade you liked.
- Ancient Olive (6002-3A): Another good match with a slightly muted tone, which could work nicely in your space.
- Secluded Woods (6002-4A): This has a deeper, earthy olive tone that could also be a good match.
These should be close to the colour we discussed. If you can get samples of these shades, it might help to see them in your room’s lighting before making a final decision.
I painted a stripe of each on my wall and took a selfie with them – this would be my Zoom background after all – for ChatGPT to analyse. It picked Secluded Woods, a charming name it had hallucinated for a paint that was actually called Brisk Olive.
I was relieved it didn’t choose the most boring shade, but when I shared this story with Jang at OpenAI, she looked mildly horrified. She compared my consulting her company’s software to asking a “random stranger down the road”.
Judith Donath, a faculty fellow at Harvard’s Berkman Klein Center, who studies our relationship with technology, said she didn’t think that using AI was much better than flipping a coin or throwing dice, even if these chatbots do have the world’s wisdom baked inside.
“You have no idea what the source is,” she said. “At some point there was a human source for the ideas there. But it’s been turned into chum.”
Each of my AI companions had a slightly different personality. Microsoft’s Copilot was overeager. Google’s Gemini was all business.
When I explained my experiment, these assistants were happy to help, with one exception: Claude, a prickly chatbot developed by Anthropic, a company worried about how AI could go terribly wrong. Claude said making decisions for me was a bad idea, and cited entirely valid concerns about the limitations of AI and how much information and control I would be handing over.
AI enthusiasts told me that Claude was the smartest of the chatbots; between that and its principled stance, I came to think of it as the Hermione Granger of the group. The others, particularly ChatGPT, were up for whatever I threw their way. When I asked Spark what haircut I should get, it recommended a textured bob, calling it “professional and polished,” and popular with celebrities. (Based on the compliments I’ve gotten since, it was a decent choice.)
Amanda Askell, a philosopher and researcher at Anthropic, told me that large language models tended to provide “the average of what everyone wants”. She wears her own white blond hair styled as a kind of mullet with baby bangs. “I don’t think AI would choose my haircut,” she noted.
Mannequin core
Halfway through the week, I found myself in a J. Crew dressing room because AI hated my clothes.
I had uploaded photos of my wardrobe to Style DNA. Based on a scan of my face, it had determined my style and optimal colour palette. Most of what I owned, including some of my favourite items, were not a good match for me, according to the AI stylist.
The app fixated on two garments – a pair of light denim shorts and a fluorescent orange exercise shirt – encouraging me to incorporate them into almost every outfit.
While I couldn’t bring myself to pair the sweat-wicking shirt with a black blazer for the office, I did appreciate the AI’s bluntness.
At J. Crew, I took mirror selfies as I tried on each item. It rejected most of what I’d selected but gave a “100% match” to an olive knit tank top and high-waisted dark jean trousers.
I uploaded the AI-approved attire to a chat with some of my stylish colleagues.
I got absolutely roasted.
“If I saw you walking down the street I would think, ‘Sure, that is a person who got the mannequin set,’ " said one. “Mannequin core,” chimed in another.
Sandra Matz, a professor at Columbia Business School, has written a coming book, Mindmasters: The Data-Driven Science of Predicting and Changing Human Behavior, about how algorithms influence what humans do. She told me that one of her biggest worries about handing agency over to AI was “the potential to make us super boring”.
There was one part of my life AI failed at: being me.
Cloning my voice was disturbingly easy to do. I also created a video avatar of myself by reading a short script into my laptop camera. I wanted it to post to social media and send video messages on my behalf.
But when it digested one of my articles for a TikTok video, the script was wooden and some of my movements were exaggerated in a creepy way. When I used my avatar to send a loving, AI-composed message to my mum, she was horrified. “You seemed so phony!!! I thought you were mad at me!!” she replied.
The messages AI composed on my behalf were overly effusive. (“Stay awesome”? Really?) Even when they reflected my own thoughts and desires, they came across as inauthentic to others, such as when I let AI craft the message to my mother-in-law letting her know she was welcome to come over to our house.
“I was really delighted by your response and I felt so loved,” she told me, “and then it struck me that it might be AI.”
When the week came to an end, I was happy to take back control of my life. Generative AI had made decent decisions more quickly than I would have on my own and hadn’t made any terrible mistakes, but it also hadn’t wowed me.
“A lot of people think that generative AI helps us to do new things or to be innovative,” said Viktor Mayer-Schönberger, a professor at the Oxford Internet Institute who writes about harnessing technology. “The real ability that generative AI brings to the table is efficiency.”
This article originally appeared in The New York Times.
Written by: Kashmir Hill
Photographs by: Kashmir Hill
©2024 THE NEW YORK TIMES