Journal of Innovation & Industry Transformation
Vol. II. 2017 Submitted: May. 2017- Approved: Sep. 2017
Date: 2017 (Updated for 2026 Context)
Andreou, Elena
Deep Tech Diplomate | Quantum Brand Strategist
This research examines the functional and ethical limitations of early-stage Artificial Intelligence (AI) through the lens of visual recognition and algorithmic classification. At a time when AI was primarily viewed as a tool for profitability and automation, this paper explores the “semantic gap” – the disconnect between mathematical computation and human-contextual understanding. Utilizing a case study approach focused on the Clarifai API, the researcher conducted a qualitative evaluation of image recognition accuracy. Using a controlled set of personal and professional imagery, the study tested the algorithm’s ability to interpret complex human concepts such as “joy,” “environment,” and “professionalism”. The findings were analyzed using Roger’s Diffusion of Innovations and Buckland’s “Information-as-Thing” frameworks.
Findings: The study reveals that 2017-era AI functioned as a rigid classifier rather than a reasoning agent. Key findings include:
Contextual Misalignment: AI consistently misinterpreted social cues (e.g., mistaking a paintball game for war).
Algorithmic Bias: The system relied on reductive physical markers (e.g., exposed knees) to assign subjective tags like “sexy,” failing to perceive the artistic or professional intent of the imagery.
Information Limitations: The research confirms that AI performance is fundamentally capped by the “Information-as-Knowledge” inserted by human programmers, leading to a “vicious cycle” of biased decision-making.
Conclusion: While AI significantly boosts productivity and data analysis, this research concludes that AI (in its 2017 form) cannot replace human judgment or sensory-based decision-making. The paper advocates for a “Human-Centered” approach to AI development, emphasizing transparency and the disclosure of algorithmic logic as essential ethical standards for the future of the digital and quantum economies.
Originally published in 2017, this paper has been updated in 2026 to evaluate these findings against modern multimodal foundation models and the current regulatory landscape of the EU AI Act.
Nowadays, all big companies and most small businesses are focused on increasing profitability and improving competitiveness. With this goal in mind, many have turned to Artificial Intelligence (AI) to automate tasks previously performed by humans. AI is receiving significant attention, and the debate is growing: will it change the world for better or for worse? This remains one of the most difficult questions of our time, with no single, simple answer.
AI is often considered the innovative technology that will replace much human labor within the next few years. Consequently, academics and researchers are deeply focused on this phenomenon. This research centers on Clarifai – an AI-based tool marketed as “Artificial Intelligence with Vision” – as a case study to explore the benefits, challenges, and ethical issues surrounding visual recognition technology.
AI involves computer systems developed to learn and perform tasks normally requiring human intelligence, such as visual perception, speech recognition, decision-making, and translation. In 2010, Singh described AI as “intelligent thought that can be regarded as a form of computation.” It is no longer a science-fiction scenario; AI is here to stay. Even those who resist it can hardly avoid the “embedded microchips and invisible processors” (Brown and Duguid, 2000, p. 74) that make our lives more predictable and reliable.
As Gleick (2011) mentioned, “every new medium transforms the nature of human thoughts.” It is normal for such innovations to provoke structural, cultural, and ethical challenges. Brown and Duguid identified the “6D’s” – radical societal changes believed to be brought about by information technology: demassification, decentralization, denationalization, despacialization, disintermediation, and disaggregation. While these changes were anticipated worldwide, time has shown that the reality of these shifts is often more complex and less uniform than initially predicted.
According to Rogers’ (2003) “Diffusion of Innovations” theory, an innovation must pass through a five-step process to be adopted: Knowledge, Persuasion, Decision, Implementation, and Confirmation. In the context of AI, a central debate remains: Human Brain vs. AI.
Most advocates state that AI will never completely replace the human brain. While early theories suggested AI was a simple algorithm, modern applications often use multiple, highly complex algorithms working in tandem. However, the core principle remains: an algorithm follows a set of rules to process input data and generally cannot produce new labels beyond those established in its model.
Furthermore, while humans use senses and “what if” questioning to make decisions, an algorithm replicates the result that the system matches as the best fit based on its programming. While some modern algorithms can learn in limited ways, they still lack the holistic “wandering” nature of the human mind.
Clarifai is an artificial intelligence company founded in 2013 that excels in visual recognition, solving real-world problems for businesses and developers alike. It has been a market leader since, winning many competitions in image classification. Their website claims that their powerful image and video recognition technology are built on the most advanced machine learning systems and made easily accessible by a clean API , empowering developers all over the world to build a new generation of intelligent applications.
Clarify calls its API: Artificial intelligence with vision! The first thing that is to be noticed is that the “ai” in different color used on their logo visually separates it from the letters “clarif”.
The logo, the moto as well as their website statement that core model identifies 11,000+ general concepts like objects, ideas, and emotions pass the message that they have managed to overcome all these barriers and differences between human brain and AI. But did they really make it?
2./
In order to better understand Clarifai’s functioning, as well as its strengths and weaknesses, an evaluation was performed on a small set of photos.
For the purpose of this research, and trying to avoid any ethical issues or any claims of intellectual properties rights infringement, the photos chosen are personal photos. All of them give a special message which a human brain can perceive easily by using its logical and its sense and the purpose is to validate as to what degree Clarifai’s algorithm retrieves the same categories.
Clarifai has 7 models for different kind of images and one for videos. For this research the general model was used.
Img. 01 and 02 “Protecting our Planet starts with you” where designs created for a fashion show project with an environmental scope and were chosen because they give a particular environmental message.
The primary design utilized iconography of water and avian life to evoke a prosperous ecosystem. In contrast, the secondary design presented a stark visual warning of ecological collapse, depicting arid, lifeless terrain – a clear representation of the consequences of current environmental neglect.
The experimental results indicate a significant failure in the algorithm’s ability to decode these higher-level semiotic cues. While a human observer immediately synthesizes the imagery of fleeing birds and parched earth as “environmental crisis,” the Clarifai system failed to recognize these specific biological and geological markers.
Instead, the program generated consumer-centric metadata such as “beautiful,” “bag,” “shopping,” and “business.” These tags suggest the algorithm prioritized the medium (clothing/commerce) over the message (environmental advocacy). Furthermore, there is no technical documentation to suggest that the API possesses Optical Character Recognition (OCR) capabilities sufficient to interpret the explicit textual warning: “Protecting our planet starts with you!”
Crucially, the system’s attribution of the category “sexy” highlights a fundamental flaw in algorithmic “Ground Truth.” The tag appears to be triggered by the mere exposure of a subject’s knees, demonstrating a reductive, pixel-based logic that ignores the thematic context of the image. This finding underscores the inherent subjectivity of machine-generated labels. Tags such as “beautiful” or “sexy” are culturally contingent and lack a universal consensus; their inclusion in an AI model introduces a layer of automated bias that fails to account for the diverse socio-cultural interpretations of the human experience.
Img. 1 (left); Img 2 (right)
Img. 03 “Paintball game” gives, at a first glance, the impression of group soldiers after a battle. But a human brain will easily recognize that these are paintball fake guns, the smiling faces and the joy of the people as well as the woman and kid show that this is a game for fun. It has to be noted that the retrieved categories by the algorithm were: “war”, “military”, “soldier” and ”weapons”, categories all related to war than “game” or “joy”. Moreover, the system tagged the category “religion” which demonstrates that the algorithm associated the image with Muslims.
Img. 03. Paintball game
Img. 04 “Celebration” photo was chosen because the message is clear and “celebration” category should have appeared. The champagne glasses, the “cheers” and festive atmosphere sent a clear message. It is to be noted that “recreation” and “restaurant” were retrieved, but nothing to imply the “celebration” act or at least describing it as depicting a romantic dinner.
Img. 4. Celebration
3. /
Img. 05 “Evening out after conference” was taken after a work event. The lady wears a white professional suit. This image doesn’t convey a particular message but it was chosen because the pose is one of a model in a photoshoot. As expected, although the photo is nothing about modeling and fashion, the algorithm because of the pose and silhouette retrieved categories as “model” “fashion”, “fashionable”. Furthermore, the system associated the image with the category “musician” which it has nothing to do with this photo.
Img. 5 Evening our after the conference
Img. 06 “FASHION-essay’s photoshoot” is analysed as expected. The categories are correct, such that prompts you to think “fashion”. From this image it can be assessed that the information inserted on the algorithm is such that when it sees short dresses or the knees one of the categories that will appear will be “sexy”.
Img. 6 Fashion-Essay photoshoot
Img. 06, Img. 07 and 08 were chosen to check the ability of the system to proper evaluate a certain category because they were evaluated by the majority of the people, fashion commentators and marketers as “sexy”.
Img. 07 “FASHION-essay photoshoot” was evaluated by the marketing team as “too sexy to be proper for the website”. As it can be noticed a lot of relative categories appeared but the category “sexy” is not, which validate the assumption that only when the system recognizes short dresses or the knee it will assess the image as sexy.
Img. 08 “Sexy dress” was chosen since the dress with backless with only two cross straps and it is considered according to all fashion journalists a very sexy dress. Justin Enriquez and Christine Rendon fashion journalists for dailymail.com have written many articles about “Bringing Sexy Back” in relation to backless dresses. In the fashion world and people that deal with clothing and fashion it is well recognized that backless dresses give a “sexy look”.
4./
A search through Google with the keywords “backless dresses” retrieved links where all of the “backless dresses” keywords are associated with the keyword “sexy”. (Img. 9)
Although, the image of the model is considered sexy, by all fashion journalists and all people who deal with fashion or clothing, it is clear that the system failed to recognize and tag such a category. This supports the proposed theory that Clarifai uses the feature “knees showing/exposed” when deciding whether or not the “sexy” tag is appropriate for a given photo.
Img. 9 Search throught Google for “backless dress”
Img. 10 “Tired, fed-up face” photo showing only a tired face although it was taken to remind a difficult day, and it was uploaded on a Facebook account. What was a surprise for the owner of the photo was that almost all comments on the photo were about “sexy look”. For the accuracy of the research, it was taken in account that from 32 people that commented on this photo, 28 of them had the above mentioned opinion. It is well accepted that a tag such as “sexy” is subjective. However, the general opinion of all fashion critics is that “sexy” has nothing to do with the amount of skin showed or if the dress is short and the legs are exposed. The other interesting fact is that Clarifai while it didn’t retrieve categories such as “sexy”, tagged categories such as “model”, “music”, “fashion” and “light” which have no logical meaning for a simple face.
As we all realize, the use of innovative technology is a must nowadays, and if we have to think of the “speed” as analysed by Steiner the advantages and economic benefits are enormous. The time that a person will need to file and classify manually photos are incomparable with the very few seconds needed by an algorithm. So, there are no doubts why this innovation was well received and embraced by many companies such as StyleMePretty.
In order to better understand the response of the public to Clairifai as a service, a Google search was performed for reviews of Clarifai. This retrieved many 5 out of 5 star reviews and many positive articles (Img. 11, 12 and 13) which lead to the conclusion that the innovation is already at the last stage of confirmation as described by Rogers.
Img. 11 Clarifai reviews
Img. 12 Article about Clarifai
Img. 13 Forum and discussion
5./
But what kind of information plays a role within this algorithm and makes it run at such performance? It could be stated that the “information as knowledge” that a programmer inserted made it possible for this API to load and process such volume of info, in the speed of light. But then as correctly pointed by Buckland, the information as knowledge becomes “information as thing” which is the algorithm itself. And then this algorithm gave “information as knowledge” to the enterprises that they use this API to analyze the photos . But if the information as knowledge was not totally correct it means that the information as thing that it will come out -the algorithm has some wrong roots, which leads to not accurate results!
A search through Clarifai’s website and policies for references about the criteria that the programmers insert the data and group it in labels and how the information is processed didn’t retrieve anything, besides the usual statements about how they collect information and how they make use and display the data. This aspect gives ground to the ethical dilemma raised by Turili and Floridi that a company should display data about these aspects too.
Brown and Duguid provide us with food for thought about “information constraints” and state that people are responsible for all information stored and they are the ones to decide “what it all means and why it matters.” They also stressed that “theorists talked about the humanity’s “bounded rationality” and the difficulty of making decisions in conditions of limited or imperfect information but what happens when humans insert wrong labels in algorithms and then these algorithms create information which will be used and needed on decision making? So, again a vicious cycle?
It was Ekstrom in 2015 that said that algorithms are systematic instructions created by humans, and that “behind every algorithm there is always a person, a person with a set of personal beliefs that no code can ever completely eradicate.”
Here we cannot talk about digital divide, but if we are to talk about maybe we should add the 5th category such as “lack of marketing or any other specialty skills” of the programmer when inserting data on the algorithms.
As the research showed, the system failed to recognize messages or emotions. Back in 1948 Claude Shannon said that human mind wanders around, and conceives different things day and night and that the human mind has the ability of questioning things.
The algorithm is learning to place the labels on the appropriate images. And we are not arguing here that the model does not have the right labels to use.
Steiner was one of the first to question algorithmic decision-making. Although he believes that we need to accept our algorithmic overloads, before doing so we should vigorously and transparently debate the rules they will impose. He continued saying that “the real question isn’t whether to live with algorithms, but how to live with them!”
Ekstrom also said that we need to filter the facts and value what is more important.
It is well recognized and accepted that AI is a technology that comes with challenges, such as accountability, security, technological mistrust, and the displacement of human workers. When Gleick said about information that is “the blood and the fuel, the vital principle” one question rose; how the society would be if at some point AI will be considered the blood and the fuel in any industry and it will be proved to be preferred instead of human intelligence? Will that be for better or for worse?
6./
The advantages of speed and economic benefits in AI are enormous. However, as Steiner (2012) emphasizes, AI is having an increasing influence in areas that affect our lives, raising serious ethical dilemmas. Turilli and Floridi (2009) argue that it is not enough for companies to display truthful information; they should also disclose how that information was gathered and analyzed.
The “vicious cycle” occurs when humans insert biased or limited labels into algorithms, which then create information used for critical decision-making. As Ekstrom (2015) noted, “behind every algorithm there is always a person with a set of personal beliefs.”
.
7./
AI is an innovation that helps increase productivity and improve the quality of life in many sectors. However, because information is not always easy to classify, we must accept that AI has limits. It cannot fully replace human judgment, especially in areas requiring deep cultural and emotional context. The real question is not whether to live with algorithms, but how to live with them transparently.
| Research Focus | 2017 Perspective (Original) | 2026 Perspective (Update) |
|---|---|---|
| Technology | Unimodal API (Clarifai) | Multimodal Foundation Models |
| The Gap | Mislabeling pixels (Knees = Sexy) | Reasoning/Logic Hallucinations |
| Ethics | "Vicious Cycle" of human bias | EU AI Act compliance & Transparency |
The following sections provide a longitudinal update to the 2017 study, bridging early-stage visual classification with the contemporary landscape of multimodal foundation models and the regulatory environment established by the European Union.
The technological paradigm has shifted from the unimodal classification systems of 2017 to the Multimodal Foundation Models of 2026, such as GPT-4o and Gemini 1.5. These modern architectures utilize Joint Embedding Spaces, which project text and imagery into a unified mathematical vector space.
Through Contrastive Learning (Radford et al., 2021), models minimize the distance between visual data and linguistic descriptions, allowing for a more nuanced “understanding” than the discrete labels used by the 2017 Clarifai API. Furthermore, architectures like VL-JEPA (Vision-Language Joint Embedding Predictive Architecture) predict continuous embeddings rather than simple tokens, enabling the system to focus on task-relevant semantics while abstracting away surface-level noise (LeCun, 2025).
However, this advancement has birthed a new iteration of the semantic gap: Cross-modal Hallucination. While 2017-era AI failed due to a lack of labels (e.g., mistaking paintball for war), 2026 AI may correctly identify the activity but confidently hallucinate a complex, nonexistent narrative about a specific historical reenactment, prioritizing linguistic “fluency” over visual “ground truth” (Li et al., 2023).
The 2017 call for “transparency and disclosure” has evolved from an ethical aspiration into a comprehensive legal mandate. As of August 2026, the EU AI Act is in full enforcement, fundamentally altering the transparency requirements for algorithmic systems.
Under Article 50, providers of generative AI are legally required to ensure outputs are machine-readable and detectable as synthetic. This is bolstered by the 2026 Code of Practice on Transparency, which mandates persistent indicators for AI-generated content (EU AI Act, 2024). Furthermore, for “High-Risk” systems, Article 13 requires Explainable AI (XAI) standards, ensuring that AI logic is sufficiently transparent for human interpretation. The “vicious cycle” identified in 2017 – where biased human data becomes biased “Information-as-Thing” – now requires mandatory technical documentation to disclose training data characteristics and performance limitations (European Commission, 2025).
In the decade since the original research, the conceptualization of AI has moved from a static repository of information to Agentic AI – autonomous systems that perceive, plan, and execute actions. This necessitates an update to Buckland’s (1991) “Information-as-Thing” framework to account for Agentic Workflows, where information acts as a trigger for autonomous agency.
In 2026, the emergence of Context Engineering allows for the embedding of relational meaning within data, enabling agents to reason through complex scenarios (Gartner, 2025). However, the “Agency Test” reveals a critical shift in failure points: if the 2017 model mislabeled an environmental crisis as a “shopping bag,” it was a failure of classification. If a 2026 autonomous agent makes the same error, it may automatically initiate a marketing campaign for sustainable fashion rather than alerting an NGO. The semantic gap has effectively migrated from identification to autonomous intent.
The reductive logic of 2017 (e.g., “knees equals sexy”) has been replaced by more sophisticated alignment techniques, primarily Reinforcement Learning from Human Feedback (RLHF). While this has mitigated crude errors, it has introduced subtle forms of Western-centric bias.
Research indicates that RLHF frameworks often encode the norms of the majority-culture labelers, leading to a failure in Cultural Grounding (Sharma et al., 2025). This manifests as a modern form of erasure; for example, a model may consistently generate “professional attire” as a Western-style suit while failing to recognize a Barong Tagalog or a Sari as equally professional. In 2026, the challenge is no longer just identifying pixels, but preventing the model from “hallucinating” a global monoculture through its preference-optimization objectives (PNAS Nexus, 2024).
The transition from the unimodal classifiers of 2017 to the multimodal, agentic systems of 2026 represents a monumental shift in computational power, yet the fundamental “semantic gap” remains a persistent challenge. As this research has demonstrated, the errors of 2017 – rooted in reductive pixel-based logic – have evolved into the complex hallucinations and cultural misalignments of contemporary foundation models. While 2017-era AI functioned as a rigid “Information-as-Thing,” modern AI operates as an “Agentic Agent,” moving from simple misclassification to the potential for autonomous misinterpretation of human intent.
The findings of this longitudinal study underscore three critical imperatives for the future of the digital and quantum economies:
From Identification to Intent: The focus of AI development must shift from increasing raw recognition accuracy to ensuring Contextual Grounding. As AI gains agency, the risk is no longer just a “wrong label,” but an autonomous action based on a flawed understanding of human semiotics.
Regulatory Accountability: The EU AI Act of 2026 validates the 2017 call for transparency. Technical documentation and Article 13 compliance are no longer optional “ethical dilemmas” but are now the legal floor for AI deployment.
Human-Centered Design: Despite the move toward “Human-Level AI,” this research concludes that AI cannot yet replicate the “wandering” nature of human sensory judgment. The “vicious cycle” of human bias can only be broken by diversifying the RLHF (Reinforcement Learning from Human Feedback) process and adopting a “Human-in-the-Loop” architecture that values cultural nuance over mathematical fluency.
Ultimately, the question raised in 2017 – “How smart can an AI get?” – is being answered not by the AI’s ability to count pixels, but by our ability to govern its reasoning. As we move further into the quantum era, the survival of human-centered decision-making depends on our commitment to transparency, cultural grounding, and the persistent recognition that algorithms, however complex, remain a reflection of the human “Information-as-Knowledge” that birthed them.
Brown, J.S. and Duguid, P. (2000). The Social Life of Information. Harvard Business School Press.
Buckland, M.K. (1991). “Information as Thing.” Journal of the American Society for Information Science, 42(5), 351-360.
Clarifai. (n.d.). Technology and Vision Models. Retrieved from https://clarifai.com/technology
Clarifai. (n.d.). Privacy Policy and Data Usage. Retrieved from https://clarifai.com/privacy
Ekstrom, A. (2015). The moral bias behind your search results. TED. https://www.ted.com/talks/andreas_ekstrom_the_moral_bias_behind_your_search_results
Enriquez, J. and Rendon, Ch. (2016). “Bringing Sexy Back!” Daily Mail. http://www.dailymail.co.uk/tvshowbiz/article-3595817/Jenna-Dewan-cuts-daring-figure-backless-dress-heads-skincare-clinic.html
European Union (2024). Regulation (EU) 2024/1689 of the European Parliament and of the Council (EU AI Act).
European Commission (2025). First Draft Code of Practice on Transparency for Generative AI Systems.
Gartner (2025). Top Strategic Technology Trends: The Era of Agentic AI.
Gleick, J. (2011). The Information: A History, A Theory, A Flood. Pantheon Books.
LeCun, Y. (2025). “From Generative AI to World Models: Joint Embedding Predictive Architecture (JEPA).” Academic Keynote.
Li, Y., et al. (2023). “Evaluating Object Hallucination in Large Vision-Language Models.” arXiv:2305.10355.
Nilsson, N. (2005). “Human-Level Artificial Intelligence? Be Serious!” Winters Publishing, 25th Anniversary Issue.
Otterbacher, J. (2017). Information in Communication Systems. Lecture Notes, Open University of Cyprus.
PNAS Nexus (2024). “Cultural bias and cultural alignment of large language models.” Oxford Academic.
Radford, A., et al. (2021). “Learning Transferable Visual Models From Natural Language Supervision (CLIP).” International Conference on Machine Learning (ICML).
Rogers, E. (2003). The Diffusion of Innovations (5th ed.). New York: Free Press.
Sharma, R., et al. (2025). “RLHF: A Comprehensive Survey for Cultural and Multimodal Alignment.” arXiv:2511.03939.
Singh, S. and Singh, S. K. (2010). “Artificial Intelligence.” International Journal of Computer Applications, 6(6).
Steiner, C. (2012). Automate This: How Algorithms Came to Rule our World. New York: Penguin Books.
Turilli, M. and Floridi, L. (2009). “The ethics of information transparency.” Ethics and Information Technology, 11, 105-112.
Author Bio:
Elena Andreou is a Deep Tech Diplomate and Quantum Brand Strategist, specializing in the intersection of emerging technology, ethics, and human-centered design. Her work focuses on bridging the “semantic gap” between algorithmic computation and socio-cultural intelligence. Originally authored in 2017 during her research at the Open University of Cyprus, this study has been expanded in 2026 to address the implications of Multimodal Foundation Models and the legal frameworks of the EU AI Act. Elena advocates for radical transparency in AI development, ensuring that the digital and quantum economies of the future remain grounded in human sensory judgment and ethical accountability.