(This article is part of Today’s Cache, The Hindu’s newsletter on emerging themes at the intersection of technology, innovation and policy. To get it in your inbox, subscribe here.)
It seems like it is the season of all things AI. ChatGPT has been in the spotlight since the start of this month. More than a million users have queried the large language model with a variety of questions. And the giant chatbot has replied to almost all queries with a Wikipedia-like ease (some of its answers were wrong though).
OpenAI’s bot was an upgrade of its predecessor GPT-3, which was launched two years ago, and it is now exclusively licensed to Microsoft. It was the first language model to show that AI could convincingly write like a human. This new bot, based on GPT-3.5, has the ability to blend text and computer codes. That means it not only replies to general text questions and answers, but also fixes programming codes (incorrectly in some instances).
GPT-3.5 and its predecessor are trained on sentences, words, parts of a speech, loads of data from social media posts, Wikipedia entries and news articles. Last year, Sam Altman’s OpenAI revealed another AI use case with the Dall-E app. The 12-billion parameter version of machine learning model could convert text to images. ChatGPT and Dall-E are what happens when Big Data meets super-fast computation.
While ChatGPT is purely a text-based interface, Dall-E uses text input to process an image-based output. These bots are one end of the AI spectrum. They are built keeping conversations between machines and humans at the centre.
At the other end is the AI-based app Lensa AI. It uses a machine learning model, Stable Diffusion, to create artistic portrait images based on the input of two dozen selfie photographs. But, the ML model uses digital images scraped from the web to process the artistic output. It is a well-equipped greenroom, but uses tools of other creators to make the user look good digitally.
The app is built by Prisma Labs. Six years ago, the company became popular when it let users transform their selfies into pictures in the style of famous artists. Lensa AI is now built on an image-based deep learning model that uses massive amount of digital art scraped from the web and LAION-5B database to train its machine learning engine.
Artists, whose works are scraped by this engine, are unable to either opt-in or -out of this algorithm. This has been the case even when Stable Diffusion was first introduced earlier this year. Some artists had raised concerns that their images were being used by the algorithm without giving them credit or payment. Their complaint didn’t end there, artists also felt threatened that such an ML model could threaten their livelihoods as deep neural networks could create digital art at a much faster rate and at scale compared to what human artists can.
Such complaints haven’t stopped Prisma from building Lensa AI for general use. Within weeks after its launch, the app took the top spot in Apple’s App Store. An annual paid version costs Rs.2,499 in India. But the company offers a seven-day free trial to play with the app. The app suggests users submit 10-20 selfies for best results. The images must be close-up shots of the user’s face, and they must have a variety of backgrounds, facial expressions, and angles.
Once users submit the images, small concentric circles zoom in and out against the app’s black background. After processing the images, the viral photo-editing app creates “magic avatars” using the subject’s facial features.
In 2018, Lensa launched its first photo editing tool. And now, this recent launch with the magic avatars feature helps transform digital self-portraits into artistic images in a variety of themes. The app is a great tool for editing photographs. And the free version has a three-image limit per day.
The photo editor has some privacy concerns even though its policy notes it doesn’t use uploaded photos for anything other than to apply filters and effects. However, it says it may use information to train its algorithms to make the app more effective. It also mentions that it will collect data to personalise content.
And when a user signs the app’s terms of use, they will be allowing Lensa AI a “perpetual, irrevocable, nonexclusive, royalty-free, worldwide, fully-paid, transferable, sub-licensable license to use, reproduce, modify, distribute, create derivative works of your User Content, without any additional compensation to you and always subject to your additional explicit consent for such use where required by applicable law.”
Now, if you are a digital artist who plans to use get some greenroom touch up with Lensa AI for your art work, those terms will come back to bite as you will be granting the app an irrevocable right to your content.
