Seeing anew: Mechanisms, outcomes, and challenges in adapting to restored sight via visual prostheses

1.1.1. Operational definition of stages

To ensure the framework functions as an analytical tool, each stage is defined by behavioral markers reported in the literature, with stage assignment determined by reported outcomes. Each stage is introduced below with additional detail provided in Table 1, defined using criteria that can be employed to classify outcomes directly from future reported results as they become available.

• Stage 1 (S1), Device Activation, is defined by reliable elicitation of percepts following stimulation. Evidence includes detection of flashes, phosphenes, or categorical light localization (e.g., quadrant-based) significantly above chance during controlled testing.

• Stage 2 (S2), Perceptual Calibration, is defined by consistent mapping between percept and stimulus properties. Evidence includes discrimination of position, motion direction, orientation, or shape under controlled laboratory conditions.

• Stage 3 (S3), Cognitive Integration, is defined by goal-directed task use of vision requiring interpretation rather than detection. Evidence includes object identification, navigation decisions, or task completion using vision as an active information source.

• Stage 4 (S4), Behavioral Generalization, is defined by transfer of visual use outside controlled testing. Evidence includes real-world mobility, activities of daily living, or spontaneous use of vision in unconstrained environments.

• Stage 5 (S5), Psychosocial Adjustment, is defined by sustained voluntary adoption. Evidence includes continued everyday use, preference for device-mediated vision over non-visual strategies, or integration into lifestyle and identity.

The stages are ordered by hierarchical dependence rather than by temporal sequence: each requires the capabilities of those below it, but progression need not be monotonic, uniformly paced, or irreversible. A stage assignment therefore describes the highest stage supported by the available evidence rather than a fixed position occupied at all times.

We will use these stages to assess the underlying capabilities of a given device by looking at population data, but they can also be used to gauge the progression of an individual patient. For the devices we will review, a stage is considered supported when the framework criteria are met by a majority of recipients across independent tasks. Progression of individual cases, including exceptional high-performing or highly motivated ones, will be discussed descriptively to illustrate variability and adaptation trajectories and to separate fundamental technological capability from user-specific variability.

Evidence for a device will be considered sufficient for a given Stage only when reported population performance is statistically significant, or there is statistically significant performance for a majority of subjects, as compared to baseline or chance. It should be noted that statistically significant performance does not necessarily imply performance levels that are meaningful or useful, as will be discussed throughout. In general, if a device provides performance at a given stage only by statistical significance against blindness, and not through performance comparable to normal vision, the device often does not qualify at the next stage.

For safety-critical tasks (such as in Stage 4), a higher reliability threshold is required. Performance must demonstrate consistently high success rates (at least 99% correct or the equivalent) without external assistance, reflecting the lower tolerance for error inherent in such behaviors.

Stage assignment is referenced to comparator-based performance rather than to absolute values of any single metric. The metrics at each stage are often reported in incommensurable units across devices, and their magnitudes are bounded by device factors such as electrode pitch and pattern coverage. A threshold placed on such metrics would therefore emphasize device hardware rather than recipient adaptation, but a recipient who identifies letters through a coarse array has performed the same act of interpretation as one who identifies letters through a fine array. That said, where an absolute criterion is needed because of potential impact, as with safety-critical behavior at Stage 4, it is included.

When multiple tasks within a given stage are reported for a device, the criterion must be met in a majority of independent assessments. This breadth requirement prevents stage assignment based on isolated success in a single narrow or highly trained paradigm. If only a single task is reported with qualifying performance, then stage classification is considered to be tentative, limited, or partial.

1.1.2. Scope and selection of evidence

This review is intended as an integrative synthesis rather than a systematic review. The goal is to examine the extent to which restoration of visual function through visual prosthesis devices is plausible by evaluating reports describing behavioral outcomes following patterned visual input to the nervous system. Because the proposed framework concerns adaptation rather than device engineering, evidence was drawn primarily from long-term clinical prosthesis studies in human participants, supplemented by cases of non-prosthetic sight restoration that produce comparable perceptual challenges. Cochlear implants are considered as a comparative model of sensory restoration, and sensory substitution devices examined in controlled laboratory settings or extended training paradigms are included where they inform mechanisms of adaptation. Studies were selected based on their relevance to understanding adaptation to artificial vision and the availability of post-implantation behavioral, experiential, or psychosocial data.

For each project described below, all reports available through indexed searches were reviewed. Publications on fully developed devices were selected over developmental, proof-of-concept, or exploratory reports. The assembled evidence is organized into a classification that describes the highest demonstrated stage within the framework. Evidence using digital or optical magnification was not included. Quantitative psychophysical measurements and structured functional tasks were considered primary evidence. Observational functional reports were treated as supporting evidence.

Psychosocial and qualitative sources were included when they provided first-person accounts, structured interviews, or validated questionnaires addressing user experience, expectations, identity, quality of life, or long-term device use. Detailed first-person patient descriptions were included along with primary scientific reports. Qualitative interviews and patient narratives were used to characterize patterns of use, expectations, and abandonment, but were not treated as prevalence estimates. The availability of systematic qualitative assessment remains limited across artificial vision technologies, with many (but not all) reports based on small samples, single-case studies, or retrospective accounts. Psychosocial outcomes are therefore discussed to identify recurring themes and adaptation challenges rather than to estimate prevalence or generalizability.

1.1.3. Selection of devices

Within the extent of visual prosthesis projects, we reviewed the ones with results from human trials that included evidence for adaptation at the various stages described in the framework, where devices are currently under development, as well as key historical projects. Projects of historical importance that are reviewed include the Argus II, Alpha IMS/AMS, and Dobelle prosthesis. We did not include Brindley and Lewin's seminal experiments (Brindley and Lewin, 1968); although foundational to the field, they were exploratory in nature, rather than the testing of a device ready for clinical deployment. Contemporary devices that are reviewed include the BVT suprachoroidal prosthesis, IMIE 256, CORTIVIS, PRIMA, and Orion.

2.3.1. Argus II

The Argus II retinal prosthesis (Humayun et al., 2012) is among the most extensively implanted visual prostheses, with nearly 400 recipients worldwide (Nadolskis et al., 2024), and it has been the subject of the largest body of literature on functional outcomes. While Second Sight Medical Products (Sylmar, California, USA; reorganized in 2022 as Cortigent) ceased development and production of the Argus II in early 2020, follow-up of implanted patients continues (Güven et al., 2023). The system uses an epiretinal electrode array driven by a camera mounted on external goggles, and is steered by head scanning rather than eye motion. Laboratory and clinical evaluations of the Argus II demonstrate reliable progression through the early stages of adaptation, while progress into later stages is uneven. We will use the Argus II as the first application of the five-stage framework; reviews of other devices will be less detailed, but summary tables of the results for each will be given to aid the reader.

2.3.2. Alpha IMS, AMS

The Alpha IMS and its successor, the Alpha AMS, were subretinal prostheses, developed by Retina Implant AG (Zrenner et al., 2011; Table 4). The devices used an approach in which photodiode-based stimulation elements were implanted beneath the retina to target bipolar cells, operating as man-made photoreceptors to functionally replace diseased natural ones (Stingl et al., 2015). The Alpha IMS incorporated a 1500-electrode array with integrated photodetection, requiring external power and control (Stingl et al., 2013), and relying on surviving retinal circuitry to convey stimulation patterns to the visual brain, but not requiring a separate camera. The Alpha AMS addressed stability and longevity concerns by replacing the IMS's analog amplifiers with digital control, and by incorporating more robust materials (Stingl et al., 2017). Although Retina Implant AG (Reutlingen, Germany) ceased operations in 2019, reports have been published evaluating device efficiency and offering important information on functional outcomes that allow us to understand how well the device was integrated into the lives of recipients.

Early reports with the Alpha IMS had positive outcomes. In a pilot study, recipients reported that with 4–6 h of daily practice over several weeks they could control eye position, improve visuomotor coordination, and perform basic tasks such as pointing to objects and recognizing simple shapes (Zrenner et al., 2011; Stingl et al., 2013). 82% (9/11) of subjects reported light perception mediated by the implant, and 64% (7/11) were able to localize a light source significantly above chance. Five subjects could accurately report the motion of dot patterns up to 35° /sec. Grating acuity was measured in six subjects with a maximum of 3.3 cpd. Geometric shapes and tableware objects presented in high-contrast settings were counted, identified, and located all significantly better with the device on than off across the population (each test, p < 0.05). Subjective feedback from recipients included acclimatization to the flickering percepts presented by the implants and successful steering of gaze and attention (Zrenner et al., 2011), as well as the ability to recognize objects and structures in non-laboratory interior and exterior settings (Stingl et al., 2013). These results are consistent with success into S3, with anecdotal support for S4.

However, longitudinal studies indicated that these gains were not reliably sustained. In a cohort of 29 patients followed for 1–6 months, visual function as measured by BaLM tests showed inconsistent outcomes. Although 50–55% of participants initially performed better with the device on versus off, scores fell back to 40% by month three, with no significant difference by month six (Stingl et al., 2015). Tasks such as grayscale perception and object recognition initially improved but reverted to system-off levels after several months, although there was often an increase in system-off performance due to a hypothesized up-regulation of growth factors (Schatz et al., 2011) that are also seen in other systems (Chow et al., 2004; Pardue et al., 2005). The maximum visual acuity (20/546 Snellen, 1.44 logMAR) was one of the highest scores reported among clinically tested devices, but the BaLM grating acuity task showed inconsistent results. The percentage of patients successfully passing the different tests varied considerably through the six months of study, with no significant improvement found for the system on versus off at the six-month mark due in part to improved system-off scores (Stingl et al., 2015), but overall results suggest difficulty in consolidating perception for Behavioral Generalization (S4). Of the 29 participants, 25 reached primary endpoints related to mobility and activities of daily living, and 21 reached secondary goals including improvements in light perception (S1) and visual acuity (S3), yet only 13 reported vision adequate for everyday use (S4) (Stingl et al., 2015). Technical constraints, including underestimated spreading of electrical fields from the 1500 photodiodes and disease-driven retinal remodeling, may have reduced the effective resolution, affecting the capability for fluent perception (Shivdasani et al., 2014).

The revised architecture, Alpha AMS, introduced refinements aimed at greater stability and durability. In a one-year study (Edwards et al., 2018), five of six subjects passed BaLM light perception and localization tests (S1, S2); the sixth subject's implant was damaged during implantation, rendering it non-functional. The same five were significantly better at a six-level visual contrast test with the device on than off (across contrasts, 62 ± 13% versus 5.7 ± 5.1%, mean and SD, p < 0.001, paired t-test; S1), and could count high-contrast tabletop objects better with the device on versus off (88 ± 16% versus 7.5 ± 17%, mean and SD, p < 0.001, paired t-test; S2). All five passed the BaLM grating acuity test at 0.1 cpd (S2), and two at 0.33 cpd. Two subjects achieved object identification by month three (S3, but not qualifying), and three subjects were significantly better at reading the time from an analog clock (S3). None passed BaLM motion detection tests. No quantitative tests were performed that would be considered for S4.

The Turano-based Independent Mobility Questionnaire evaluating the difficulty of performing various activities showed substantial improvement in confidence with walking in familiar areas and avoiding obstacles (Edwards et al., 2018). Subjective experiences included the ability to locate balls on a billiard table, see windows and doors, kitchen appliances, high-contrast laundry, buildings, a mountain against the sun, a car, and Christmas lights. Despite none of the subjects achieving independent mobility outside their home, the broad-based questionnaire improvements indicate a high level of adaptation and satisfaction, suggesting Psychosocial Adjustment (Stage 5) is possible.

In an attempt to improve outcomes from these earlier patients, an intensive rehabilitation plan was explored for one patient three years after implantation (Kapetanovic et al., 2020). This patient had reported a benefit at 12-months, but their motivation had subsequently dropped, finding that the use of the device became an impediment, especially in familiar environments. The second round of training included screen-reading, complex shape recognition, eye-hand coordination, path following, obstacle recognition and outdoor distance estimation. Daily 2-h sessions for five days allowed the subject to successfully recognize small words and shapes (1.39 and 2.2 logMAR respectively), to identify obstacles and estimate distances in daytime and nighttime conditions, and had re-established eagerness for the implant (Kapetanovic et al., 2020). These qualitative results suggest that the patient had progressed into S4, and thus that post-implantation training has an important role to sustain or restore user engagement during plateaus in perceived benefit.

2.3.3. The Dobelle prosthesis

The first camera-based visual prosthesis implanted in a human subject was the Dobelle cortical prosthesis (Dobelle et al., 1979; see also Brindley and Lewin, 1968 for seminal proof-of-concept experiments, and Tassicker, 1956 with a design that anticipated contemporary direct imaging approaches such as currently used by PRIMA). The system used a glasses-mounted camera whose output was converted into electrical stimulation that was delivered to an array of surface electrodes on the occipital cortex. Many of the scientific publications on the Dobelle prosthesis are either qualitative, or case reports, although there is one retrospective study that will be included below.

Early reports described that recipients perceived multiple phosphenes (Dobelle et al., 1974) and could, after training, perform visually guided behaviors such as navigating around obstacles or locating objects within a room (Dobelle, 2000). One participant, known as Jerry, reportedly learned to operate the device within a day and demonstrated task-directed mobility.

However, functional use remained limited. Jerry used the system only a few hours per day in a structured routine, and another recipient, Jens Naumann, described restricted use during laboratory sessions and eventual abandonment after persistent frustration and device failure (Naumann, 2012). Even in these early demonstrations, performance depended on deliberate scanning strategies and sustained cognitive effort rather than spontaneous perception.

A later retrospective cohort study interviewing thirteen of the sixteen participants from the second phase of the Dobelle program, which used camera-based implants, provides a broader account of recipient experience (Lane et al., 2015). Participants consistently reported perception of phosphenes varying in number, color, and location, confirming reliable elicitation of visual percepts. Some individuals used these percepts for orientation, mobility, and object detection, yet functional ability did not correlate with phosphene count and many participants were unable to organize percepts into meaningful patterns. Functional interpretation therefore varied widely across individuals and across time within the same individual. Reported performance depended heavily on concentration and strategy, and perception often failed to generalize beyond specific trained situations.

The same cohort described strong emotional responses including anxiety, frustration, resentment, and a sense of loss when percepts disappeared or failed to meet expectations (Lane et al., 2015). This idea of fear of experiencing loss of vision for a second time was echoed by potential prosthesis recipients in a different study (Karadima et al., 2023). Dobelle participants frequently reported a mismatch between anticipated visual restoration and the actual performance. Engagement was often motivated by hope or participation in pioneering research rather than sustained functional benefit. These reports parallel earlier individual accounts describing enthusiasm after implantation followed by reduced use when everyday benefit remained limited (Dobelle, 2000; Naumann, 2012).

These qualitative reports suggest that the Dobelle system reliably generated percepts and occasionally supported task-directed visual behaviors, but interpretation of the percepts remained inconsistent and did not generalize into independent everyday reliance. Percept generation and intermittent goal-directed use were demonstrated without evidence of stable behavioral generalization or durable adoption. In combination, these observations suggest an assignment of S3 that is tentative due to lack of reported methodological detail (Table 5).

2.4.1. Suprachoroidal visual prosthesis (BVT/Phoenix 99)

The BVT suprachoroidal visual prosthetic device (Ayton et al., 2014) uses an electrode array placed in the suprachoroidal space between the sclera and choroid of the eye. It functions much like an epiretinal prosthesis in that the imaging camera is decoupled from the array of stimulating electrodes, and, like the Argus II, lacks gaze contingency, so subjects must be trained to hold their eyes fixed in the orbit and scan with head movements alone. Studies assessing psychophysical performance, mobility and daily living outcomes with respect to device usage suggest a lack of integration into normal activities (Petoe et al., 2021, 2024; Karapanos et al., 2021; Ayton et al., 2014). In a pilot clinical trial of the first-generation, 33-electrode implant (Ayton et al., 2014), three patients were followed for 12 months and evaluated with the BaLM (Bach et al., 2010) and Landolt C module of FrACT tests (Bach, 1996). Reliable phosphenes were elicited from all three subjects (Sinclair et al., 2016). Although participants generally benefited from the use of the system, the improvements were modest: BaLM light localization with the device on was significantly above chance and above performance with the device off (P1, 98% vs 28%; P2, 71% vs 25%; P3, 67% vs 25%), yet the mean Landolt C visual acuity (P1, logMAR 2.62, Snellen 20/8397) was below theoretical prediction for the electrode array (logMAR 2.33, Snellen 20/4242) and identifying optotype orientation in the task required approximately 1 min of time (Ayton et al., 2014).

Subsequent studies with the second-generation, 44-electrode BVT device assessed performance on quantitative laboratory tasks as well as quality-of-life instruments (Petoe et al., 2021, 2024; Karapanos et al., 2021; Titchener et al., 2020). Screen-based square localization tasks showed significantly better performance with the device on versus off for all subjects (p < 0.001; 9.5–11.4° with the device on versus 22.0–36.0° with the device off, and 13.8–15.7° with the electrode map scrambled). Two of four subjects performed a 4-way motion discrimination task significantly better with the device on versus off at all tested speeds, up to 30° per second. No subject was able to consistently meet the passing criterion in a grating detection task. All four subjects were significantly better at locating high-contrast objects on a table (p < 0.01), but only two were significantly better at identifying those objects (both p < 0.05). These test results show stage S1 for the population, and stage S2 for two of the four individuals.

Three of four subjects were significantly better at a laboratory-based door location task with the device on than off (p < 0.001), but none approached the safety-related S4 thresholds. The same three were significantly better at a grasping task that required visuo-motor coordination (p < 0.001), but all had significantly higher accuracy for unsuccessful misses (p < 0.01).

Gradual improvements on the FLORA instrument were seen. Participants described that finding doorways and locating familiar objects became progressively easier with the device on, shifting from difficult to moderate, while tasks remained mostly difficult with the system off (Petoe et al., 2021). By the study's end, mobility was moderate on average (Petoe et al., 2024). Participants reported experiences such as detecting obstacles, sorting objects, estimating distances, and counting people, with most of the quantitative improvements during the period 4–10 months after implantation (Petoe et al., 2021; Karapanos et al., 2021). These reports suggest intermittent progression into Cognitive Integration (Stage 3) and attempts at Behavioral Generalization (Stage 4).

Device use outside the lab, however, was limited. According to the device logs, the frequency of use was every two to three days for 2.3 h each time for two subjects (Petoe et al., 2021), later decreasing to once every two weeks, especially when COVID-affected periods eliminated social life and mobility outside (Petoe et al., 2024), and for six 30-min occasions in total for one subject (Petoe et al., 2021). Recipients described that after initial exploration to determine where and when their device would be useful, they used it in specific circumstances, often only for 2–3 h at a time (Petoe et al., 2024). These usage patterns of brief engagement indicate that most patients did not advance to generalized use, and thus evidence does not yet exist for S4 or S5.

A third-generation, 99-electrode device is under active development as the Phoenix 99 (Eggenberger et al., 2021). With publications from clinical testing of the new design, or extended long-term testing of the 44-contact version, the stage classification (Table 6) may change.

2.4.2. IMIE 256 epiretinal array

In five subjects with end-stage retinitis pigmentosa, the 256-channel IMIE epiretinal prosthesis produced reliable phosphene perception and measurable task performance over a 90-day rehabilitation program (Xu et al., 2021). Training consisted of repeated high-contrast laboratory exercises paired with structured mobility activities, while acuity tasks were held back from the training regimen to limit practice effects. With the system on, all subjects achieved the single tested grating acuity (20/800) and Tumbling-E acuity (20/1200), detected motion direction, localized targets, and performed orientation-and-mobility tasks such as door finding and line following with high accuracy, whereas device-off performance was absent or near chance. The limited numbers of trials dilutes on-vs-off statistical power, making the comparisons not significant despite the positive reported performances. Because testing focused on these practiced tasks and did not vary aspects like stimulus difficulty or contextual setting, it remains unclear whether performance reflected general visual capability or learned task routines. Subjects had unlimited time on each task, with typical response times in the 1–3 min range at the end of training.

These results suggest that intensive, structured rehabilitation enabled recipients to move through Device Activation and Perceptual Calibration (Stages 1 and 2) and into Cognitive Integration (Stage 3), with some evidence of early Behavioral Generalization (Stage 4) in simulated daily-living activities (Table 7). Although the subject pool was small (n = 5), the consistency of improvement underscores the role of training in facilitating adaptation. Follow-up, long-term study remains necessary to determine if these gains can be sustained and expanded into complex real-world contexts, translating ultimately into Psychosocial Adjustment (Stage 5).

2.4.3. CORTIVIS cortical prosthesis

The CORTIVIS cortical prosthesis uses a Utah intracortical microelectrode array implanted in primary visual cortex to deliver visual information. Initial published results described one participant followed over approximately six months (Fernández et al., 2021), and subsequent reports from the same clinical program indicate additional implanted participants, with at least four participants tested by 2026 (Grani et al., 2025; Lozano et al., 2026; Alfaro et al., 2026). During an initial two-month period of daily 10–15 min training sessions, the first participant learned to reliably localize electrically evoked phosphenes and distinguish them from spontaneous percepts, establishing Device Activation and enabling Perceptual Calibration (S1–S2). Over the subsequent four months, multi-electrode and sequential stimulation patterns were introduced, allowing discrimination of oriented forms, recognition of object boundaries, and identification of specific letters, including distinctions between lower- and upper-case presentations across repeated trials, indicating continued progression into Cognitive Integration (S3). Later camera-driven stimulation produced recognizable silhouettes and shapes, but performance depended on deliberate interpretation under supervised testing rather than spontaneous visual experience. A subsequent CORTIVIS case study further showed that perceived phosphene location remained gaze-dependent, and that video-based tracking of a cosmetic eye could support a gaze-contingent screen-search task in an enucleated participant, adding to the importance of sensorimotor coupling and control design in visual prostheses (Waclawczyk et al., 2025).

Within the framework, these results demonstrate reliable percept generation and spatial mapping followed by learned pattern interpretation, while leaving open whether behaviors would generalize beyond guided experimental contexts (Table 8). No independent everyday use or autonomous navigation was reported, so Behavioral Generalization (Stage 4) was not established and Psychosocial Adjustment (S5) cannot yet be assessed. Subsequent methodological work from the same program has also reported recordings from three blind human volunteers with implanted Utah arrays, indicating that the patient cohort has expanded, although the scope of the report did not include stage-defining behavioral outcomes (Lozano et al., 2026).

2.4.4. PRIMA subretinal device

The PRIMA subretinal device (Palanker et al., 2020) integrates small photovoltaic sites with stimulating electrodes to create dense pixelated patterns of electrical stimulation in the retina. It has demonstrated sustained functionality over 4 years in feasibility studies (Muqit et al., 2024) and native prosthetic acuity closely matching the implant's theoretical resolution in a pivotal trial (1.32 ± 0.16 logMAR at 12 months; Holz et al., 2025; Table 9 includes additional reports with zoom use). In early feasibility cohorts, all implanted subjects showed prosthetic light perception in the central scotoma and achieved prosthetic acuity closely matching the 100 μm pixel pitch (about logMAR 1.39, Snellen 20/500) without magnification (Palanker et al., 2022). With training, participants recognized letters and sequences of letters and transferred this ability to reading words on packages and signs (Muqit et al., 2024) including at home (Holz et al., 2025). The transparent glasses allowed simultaneous use of central prosthetic vision and residual peripheral natural vision in the same eye (Palanker et al., 2022). In a multicenter trial, most participants demonstrated improved acuity relative to baseline while peripheral natural acuity remained unchanged (Holz et al., 2025). Enhanced performance could be obtained with electronic magnification, but non-zoomed acuity reflected the spatial sampling limit on the retinal surface of the implant.

Within the framework, we will consider only non-zoomed performance in order to have consistent comparisons with other projects (Table 9). PRIMA reliably achieves Device Activation and Perceptual Calibration (S1–S2) through stable central percepts and reproducible spatial form perception. Recognition of letters and words demonstrates Cognitive Integration (S3). Although some transfer to real-world printed material was reported, the absence of statistically significant improvement in quality-of-life measures (Holz et al., 2025) suggests that consistent Behavioral Generalization (S4) and Psychosocial Adjustment (S5) have not yet been established.

2.4.5. Orion visual cortical prosthesis

The Orion Visual Cortical Prosthesis (Levy et al., 2024) adapts the Argus II architecture for direct cortical stimulation, using surface electrodes implanted over the interhemispheric portion of visual cortex. In an early feasibility trial, four of six participants provided detailed narratives after 2.8 to 3.5 years of device use that was limited to 2 h per day by safety protocol (Levy et al., 2024). All participants reported perceiving phosphenes yet emphasized that the experience fundamentally differed from natural sight, requiring months of training with research staff and low-vision therapists to interpret patterns (Levy et al., 2024; McDonald, 2019). One subject compared it to viewing a radar screen: useful for detecting contrast but lacking depth and detail. Others described the process of interpretation as akin to learning a new language (Levy et al., 2024). Similar descriptions appear in an autobiographical account of Orion use, where the percept is described as sparse flashes that must be interpreted through deliberate scanning and training rather than immediate visual recognition (McDonald, 2019). Dynamic current-steering experiments further showed that sequential stimulation could produce recognizable shapes and traced letters rather than isolated flashes (Beauchamp et al., 2020), and interim multi-subject testing demonstrated improved localization and motion-direction discrimination with the system on compared with off (Dorn, 2020). The absence of gaze contingency likely increased the difficulty of achieving stable Perceptual Calibration (S2), as its absence interferes with normal scene building (Caspi et al., 2018; Paraskevoudi and Pezaris, 2021; de Ruyter van Steveninck et al., 2024).

Despite these challenges, some participants reached Cognitive Integration (S3), using Orion for investigator-directed tasks such as sorting laundry by brightness, identifying doorways, or following the contrast between sidewalk and grass. However, functionality did not consistently translate to benefit; one recipient could detect sidewalk edges with the device, but concluded the device's utility as “marginal,” as it did not eliminate reliance on a cane (Levy et al., 2024), an important consideration for device acceptance (Karadima et al., 2023). Across reports, the device enabled certain tasks, but the inability to integrate artificial vision into real-world contexts highlighted the distinction between usability and usefulness (Table 10). Barriers to further adaptation included the cumbersome design of external components, difficulties with charging and operation, and aesthetic concerns, with some describing the headgear as unappealing and drawing unwanted attention (Levy et al., 2024), similar to the concerns expressed by Argus II patients (Erickson-Davis and Korzybska, 2021). Psychosocial Adjustment (S5) was partial: while some derived personal meaning and satisfaction from participation, and requested additional training, others found the benefits too limited to justify daily use (Levy et al., 2024; McDonald, 2019).

5.1.1. Stage 1–2: consistent achievement

Across retinal and cortical prostheses, the most reproducible findings are device activation, phosphene detection, coarse localization, and simple discrimination (da Cruz et al., 2016; Edwards et al., 2018; Petoe et al., 2024; Fernández et al., 2021). These early stages represent the most dependable achievement of current systems.

5.1.2. Stage 3: structured perceptual interpretation

Stage 3 is less common, but clearly present. Some recipients identify letters, discriminate simple forms, interpret constrained stimuli, or use device-mediated input for goal-directed action. Argus II provides the clearest evidence: closed-set letter identification improved markedly with the device on, and fine hand-movement tracing showed reduced error, although performance remained slow and effortful (da Cruz et al., 2013; Barry and Dagnelie, 2012). Similar patterns appear in subretinal systems, Alpha IMS/AMS, PRIMA, and IMIE, where some subjects read letters and words or perform controlled recognition tasks, but within narrow task structures and with substantial intersubject variability (Zrenner et al., 2011; Edwards et al., 2018; Muqit et al., 2024; Xu et al., 2021).

Comparison with other sensory interventions clarifies the significance of that pattern. Cochlear implants commonly progress beyond early ambiguity into everyday listening benefit, quality-of-life improvement, and durable incorporation into education, work, and social life (Blamey et al., 2013; McRackan et al., 2018; Spencer et al., 2012). Sensory substitution devices likewise show that the brain can learn non-canonical mappings and extract meaningful spatial information after prolonged training, although often with substantial attentional demand and limited everyday adoption (Amedi et al., 2007; Ward and Meijer, 2010; Maidenbaum et al., 2014; Chebat et al., 2018). Surgical restoration after extended blindness provides the complementary and cautionary lesson that restored retinal input does not guarantee immediate coherent visual understanding (Fine et al., 2003; Gregory and Wallace, 1963; Šikl et al., 2013; Dickens et al., 2021).

5.1.3. Stage 4–5: rare or absent

Transfer to generalized behavior with visual prostheses remains uncommon. Independent scene interpretation, open-ended object search, unscripted navigation, and sustained everyday reliance are rarely documented at the population level. Some recipients show benefit in simplified mobility tasks or selected daily activities, and some report subjective value, but broad Stage 4 support is uncommon, as is Stage 5 (Dagnelie et al., 2017; Geruschat et al., 2016; Edwards et al., 2018; Petoe et al., 2024). Patient-reported benefit is seen, but does not preclude heterogeneous daily reliance, declining observed use, fatigue, and abandonment (Geruschat et al., 2016; Erickson-Davis and Korzybska, 2021; Güven et al., 2023; Petoe et al., 2024). Empirically, most devices lie at or below the boundary between structured interpretation (S3) and generalized behavioral use (S4).

Adaptation trajectories are also heterogeneous across individuals and frequently can be non-monotonic. Gains present at one interval may be absent at another, as when Alpha IMS performance advantages had disappeared by six months (Stingl et al., 2015) and Argus II self-reported benefit declined from 80% to 65% between the first and third post-implant years (Ho et al., 2015). Interpretation can vary across time within the same individual (Lane et al., 2015), and disengagement can be reversed by renewed training (Kapetanovic et al., 2020). Variability in cortical plasticity following prolonged deprivation likely contributes: occipital cortex is recruited cross-modally in the long blind (Fernández and Merabet, 2011), yet remains available to trained non-canonical codes (Striem-Amit et al., 2012), and outcomes after surgical restoration depend on age at treatment and duration of blindness (Lewis and Maurer, 2005; May et al., 2022). Trajectory shape, rather than endpoint alone, is therefore informative, which motivates the recommendation below for longitudinal monitoring of device use.

5.2.1. Lack of skill transfer

The most consistent point of failure is the transition from trained task performance to transferable visual behavior. Skills acquired during training on one task do not consistently extend to other tasks or novel situations (da Cruz et al., 2013; Fernández et al., 2021; Petoe et al., 2024). Instead, performance commonly remains tied to the details under which it was acquired, including stimulus class, contrast regime, spatial arrangement, response options, or scanning routine. A subject may learn to identify a limited letter set, discriminate a geometric form, trace a high-contrast line, detect a doorway, or localize a salient object, yet those abilities falter under more complex environments (Dagnelie et al., 2017; Stronks and Dagnelie, 2014).

Response time carries particular importance. Many successful Stage 3 tasks require tens of seconds per trial, and sometimes multiple minutes (da Cruz et al., 2013; Ayton et al., 2014; Xu et al., 2021). That time scale is incompatible with ordinary visual behavior, which depends on rapid evaluation of scene structure to support continuous movement planning and updating. Slow performance suggests sequential problem solving through partial evidence and deliberate inference rather than fluid visual perception. That dissociation may help explain how statistically significant task gains do not necessarily imply real-world utility.

The broader comparative literature supports the same interpretation. Cochlear implant users also begin with low-fidelity, unfamiliar signals, but speech perception typically becomes faster, more accurate, and less effortful, allowing conversation in ordinary settings (Blamey et al., 2013; Ma et al., 2023). Surgical visual restoration after long deprivation often yields normal visual acuity without normal visual understanding, showing that the presence of signal alone is necessary but not sufficient for higher-order perception (Fine et al., 2003; Gregory and Wallace, 1963; Dickens et al., 2021). Sensory substitution devices can support meaningful interpretation, but users often report high attentional cost and selective use (Ward and Meijer, 2010; Maidenbaum et al., 2014; Buchs et al., 2019).

5.2.2. Possible reasons for failure of generalization

Two potential explanations recur across the literature: task-specific learning and high cognitive load. First, recipients may solve laboratory tasks through memorized stimulus-response associations rather than through formation of broadly useful visual abilities. A trained letter set, limited number of objects, or familiar tracing pattern is well suited to template matching. Such strategies can be effective for a given task, but have little utility outside the trained context, consistent with broader accounts of perceptual learning in which improvement can remain closely tied to trained stimulus conditions and task demands (Ahissar and Hochstein, 2004; Sagi, 2011). Second, interpretation often appears effortful rather than automatic. Users describe deliberate decoding, concentration, and fatigue; behaviorally, long response times and sensitivity to task difficulty point to high inferential demand (Erickson-Davis and Korzybska, 2021; Levy et al., 2024; McDonald, 2019; da Cruz et al., 2013; Ayton et al., 2014; Xu et al., 2021). Learning under those conditions is slow, and the resulting competence is narrow.

5.2.3. Contributing factors to cognitive load

Several signal-level factors likely produce cognitive load. The first is low information content. Beyond the typically coarse resolution of current visual prostheses, nominal electrode count can overstate useable spatial information because current spread, receptive-field overlap, and overlapping activation volumes can produce crosstalk or phosphene-phosphene interaction, while fading and disease-driven circuit remodeling further reduce effective channel independence (Shivdasani et al., 2014; Sinclair et al., 2016; Bosking et al., 2022; Pérez Fornos et al., 2012; Moleirinho et al., 2021; Jones et al., 2012). The resulting signal may not support the spatial resolution required for scene segmentation and object disambiguation.

The second is encoding mismatch. Electrically-evoked neural patterns differ substantially from natural retinal, thalamic, or cortical activity (Maturana et al., 2018; Millard et al., 2015). Even when downstream circuitry can extract some structure from the imposed code, the signal lacks many of the spatial and temporal regularities on which ordinary vision depends (Pérez Fornos et al., 2012; Jepson et al., 2014; Millard et al., 2015; Meital-Kfir and Pezaris, 2023). The brain may therefore learn correspondences to that code without converting it into a natural perceptual representation.

A third factor is distorted spatial representation. Phosphene position, size, and shape vary across electrodes and, in some systems, across time (Sinclair et al., 2016; Bosking et al., 2022). That inconsistency weakens the reliability on which visuospatial learning depends. If similar inputs do not reliably produce similar sensations, each perceptual judgment requires separate interpretation rather than automatic readout. A narrow field of view compounds the problem by forcing serial sampling. Users must scan over time to accumulate information that natural peripheral vision would provide in parallel, with implications for mobility and safety (Ayton et al., 2014; da Cruz et al., 2016; Petoe et al., 2024). Restricted dynamic range adds a further constraint by compressing contrast and fine spatial structure. Decoupling from the oculomotor plant imposes yet another burden: eye and head movements may fail to generate the perceptual shifts expected by the visual system, especially in head-steered systems (Caspi et al., 2018; Paraskevoudi and Pezaris, 2019; Paraskevoudi and Pezaris, 2021; de Ruyter van Steveninck et al., 2024; Waclawczyk et al., 2025). Under all those conditions, the user must learn an unfamiliar code and an unfamiliar way of sampling the world.