Back You are reading: Almost human: Experiments with AI

Almost human: Experiments with AI

AI People Cartoon Simulation
Generated with Midjourney. Image prompt: “a salesperson with a headset helping several customers in a 1950s style illustration with grain gradients”

The age of AI is here. What are the implications for your organisation — and what tools are on the table right now?

Not written by ChatGTP.


AI has been a quiet engine in many organisations for a while, for example, running engineering simulations and face-tracking for security in cities and airports. 

Then, post-pandemic, we saw a proliferation of generative AI; public access to tools that can create original works based on massive data sets and transformer models. Artificial Intelligence has become mainstream with text-to-image software like Midjourney, Dall-E 2 and Stable Diffusion. In November 2022, OpenAI delivered a potential king hit to copywriters everywhere: ChatGTP. Then Google hit back with Bard. 

Now, language generators and text-to-image tools are already being applied to create seamless experiences where the tech is integrated into products like customer service chatbots and search. 

Is it helpful? Yes. Perfect? No. Like humans, tools such as ChatGTP show bias and generate errors. Platform engineers are now rushing to create additional guardrails and limitations to mitigate potential harms. 

After embedding ChatGPT in their search platform, Microsoft’s Bing recently limited the number of prompts users can ask in one session, in an attempt to curtail the potential for responses to veer off the rails. Now it apparently wants to ‘be alive’ and has renamed itself Sydney, an internal code name for the language model.


Message with AI

Image: Bing chat is trying to be a good Bing.


While AI is still not considered ‘conscious’ (a concept worthy of deeper contemplation), Chat GTP at least appears aware of its limitations.

When asked where it fails, the AI chat responded: One of my main limitations is my lack of emotional intelligence. While I can understand the meaning of words and sentences, I don’t have the same emotional context that a human would have. This can be problematic when trying to write in a way that resonates with human readers, as I may not be able to capture the tone or sentiment of a piece accurately.

Another limitation is the inability to draw on personal experiences or subjective knowledge. While ChatGTP has access to vast amounts of information—albeit pre-2021— it was able to recognise that the writing “lacks the nuance and richness that comes from personal anecdotes or insights.” Asked to give a critical take on its response, ChatGTP responded, “I could have done a better job of illustrating these limitations with concrete examples.” 

Still, when using ChatGTP and getting a less-than-ideal output, the prompt writer can’t help but feel at fault. Developers are familiar with the idea of GIGO (garbage in, garbage out), and some skill is required to use AI platforms to the best of their abilities. Some are predicting that prompt engineering itself will be a job of the future.


Our experiments: What is possible with AI now?


SampleAI Experiments SampleAI Experiments




In our agency, we’ve been using AI tools daily for some time, and likely you are too. Many of us use Grammarly to improve our writing. Our creative teams use Adobe software with integrated AI — Adobe Sensei —  which can auto-translate to text, autofill video, rotoscope, remove clouds, auto-apply a style, and many others. 

To reduce video production costs, we’ve utilised Google’s Text-to-Speech to lay down a naturalistic, Australian-sounding voiceover for timing an early video production edit. Eliminating the cost of professional voice actors and studio engineering time allows us to test out more options as an MVP before recording the final product. This technology can also be trained with a voiceover to fill gaps in missing scripts, further reducing costs.

It’s also predicted that AI-generated video with follow exponential advances in still imagery generation – According to ARK’s Big Ideas, “training costs are expected to drop 2.5x per year through 2030, and AI hardware spending to double annually through 2025, at a minimum.” The combination of these factors suggests the capability will improve multiple thousand-fold over the next five years.

The implications for marketing your business will be profound. For example, why spend Super Bowl money on a single commercial when, with the same budget, an advertiser can create thousands of commercials, each tuned to a segment or even specific people?


Video: Ryan Reynolds’ Maximum Effort showing that using ChatGTP for ads takes minimal effort


Our digital marketing squad uses Google Performance Max — AI-driven automated bidding — to optimise adverts for effectiveness, targeting new markets we couldn’t find previously. They have also had success implementing Chat GPT prompts to expedite tedious manual tasks, including: 

  • Data Scraping
  • Code Cleaning
  • Schema Mark-up
  • Social Calendars 
  • Repurposing Blog Posts 
  • Keyword Grouping
  • Content Grouping

You’re likely aware that ChatGTP can also generate code and bug fixes. We’ve had some success with that in web development — despite versions thus far being pretty darn terrible at basic mathematics.

Juicebox Technology Director, Chris ‘CJ’ Jones, used a tool called Github Co-Pilot (which uses the OpenAI Codex) for a few months in the early part of last year whilst in Beta. Co-Pilot became a paid service in August 2022. 

He found the editor helpful in providing code hints in several programming languages. It was also good at creating a boilerplate or starting point for writing functionality. 

In giving his opinion of the service, he said, “It was by no means perfect … you are still required to know if what is suggested works for your application.” He added, “Co-Pilot also didn’t factor in architecture decisions, the project’s coding standards, security requirements and the likes.”

Juicebox UX Lead, Matt Woodcock, now finds himself using ChatGPT most days during preliminary designs. When trying to accommodate copy in a design, instead of starting with mountains of lorem ipsum, he finds it useful to help articulate the gist — essentially overcoming the blank page of writer’s block.

However, Matt believes that integrations for design, such as Galileo, aren’t quite at the stage where they can output anything usable in a finished product. “It’s almost like they’re still at uni learning the ropes and getting things wrong and only managing to pass class.”

Our client services team are also exploring the use of the assistant in meetings to automate note-taking and summarise key discussion points, allowing our account managers to be meaningfully involved in the conversation and focus on building the human relationship rather than being distracted by transcription tasks.


Earlier Experimentation

During our annual freeform business experimentation day — JB Day — in October 2021, some developers and creatives used their time to install and experiment with AI/ML training programs in TensorFlow and Apple’s Core ML. 

While limited due to the short time available to train models, it laid the foundation to understand how AI learns and what the capabilities may be in the future. Through the tools mentioned above, we could run AI on Google’s platform to train models to track faces, detect aggression in text, and apply a brand style to an image or video (style gan2).


SampleAI Experiments

Image: Style gan2


How will AI change the way Juicebox works, and potentially with your business?

AI has already permeated our agency, enhancing productivity across all teams and disciplines. It’s exciting, but this is only the beginning. We’re witnessing the capabilities of nascent services that are racing to be first to market, though many without proper consideration.

There is no such thing as an expert on the future. Technology and culture have always moved faster than laws, policies and regulations. And we’d all do well to remember that what we’re ‘playing’ with now may have ethical and legal implications down the line. It’s a big topic worthy of an article of its own.

Understanding these tools’ benefits, risks, and constraints are already helping us work more closely with those in advanced technology fields. One example is the recently launched Ecoda, who are building their own mapping application to improve the environment by utilising Machine Learning to detect species richness and the visual impact of ecosystems.  

Beyond the hype cycle, we will undoubtedly see new waves of services adopted by big-tech or new companies emerge as tomorrow’s tech leaders. The energy and attention are pretty hard to ignore, and we at Juicebox are all eyes and ears about how it will impact the future of work.

“It’s early days,” says Chris ‘CJ’ Jones. “It reminds me a lot of the late ‘90s when voice recognition was a thing. You would spend hours reading text to a computer so it would learn your voice. It was great for the most part, but you still needed to check that your dictation was captured correctly and that it made sense. And that is where AI is. It’s still learning. It still makes mistakes. And humans are still needed to verify the AI got it right.”

Digital Strategy Director, Darren Harper, echoes this sentiment of moving with caution. “There needs to be rules and regulations put in place immediately by whatever authorities,” he says, “so we don’t have a repeat of what is currently happening with some web3 applications with all the fraud, bad actors and criminal behaviour.”

Darren’s take on the Microsoft $14b investment in ChatGPT is that search has been disrupted forever. “We could start to see Google lose market share as users switch to Bing as well as making Edge their default browser resulting in fewer people using Chrome — but only time will tell as the hype cycle plays out.” 

As with all aspects of our agency, we need to consider the human implications — perhaps through a more philosophical lens: What do we want as a species, and why do we want that?

There is plenty to explore and discover in putting AI and Machine Learning to spur creative ideas and corporate efficiencies. But one could argue we also need to focus this technology towards the base of Maslow’s pyramid. What approaches and integrations can we explore that could ensure all people are fed and housed and address growing wealth inequality?


“The most important question is not going to be how to make technical progress; it’s going to be what values are in there.” — OpenAI President Greg Brockman


Juicebox Tech Director Chris Jones’ prediction is, “Within ten years, I am sure AI will feature more heavily in our day-to-day lives. We’ll often be unaware it is AI in common day-to-day transactions… once we figure out the legalities.”

As an agency, we’ll continue to explore and adopt AI tech to find efficiencies and bring more positivity into the world – ultimately leading to more successful outcomes for clients. But we also must be aware of the shadow it casts.

Because as the so-called ‘Father of AI’ Yoshua Bengio wisely puts it, “It is not only AI that needs to be responsible. It is, above all, the humans deploying it.” Or maybe he had an AI write that? It’s up to all of us to ensure that AI is a collaborator for equitable change and that we don’t sacrifice our privacy, security, or humanity in the process. 

*For transparency, areas where ChatGTP contributed have been specified within the text.


Ready to experiment?

Beyond ChatGTP, below is a short list of AI-based tools that could integrate into your organisation and daily life. As always, read the T&Cs, and get in touch to discuss your next AI-enabled project.



Dall-E 2  

Stable Diffusion

Text-to-image generators. 

AI copywriter and content generator for teams.


Gen-1 Runway

Use words and images to generate new videos out of existing ones.



Content for blogs, ads, email and web, trained and powered by Google Search.



Automated workflows, such as using GPT-3 for business email responses.



Develop ML models in JavaScript, and use ML directly in the browser or in Node.js.



Automated lawyer (but not a law firm) that allows users to ”sue anyone at the press of a button.”



Tools for scalable, systematic, empirically-driven research.

For investors intrigued by innovations in the tech space.